=Paper=
{{Paper
|id=Vol-2104/paper_201
|storemode=property
|title=Neuro-Fuzzy Model of Development Forecasting and Effective Agrarian Sector
Transformations of Ukraine
|pdfUrl=https://ceur-ws.org/Vol-2104/paper_201.pdf
|volume=Vol-2104
|authors=Diana Nemchenko,Vitaliy Kobets,Larisa Potravka
|dblpUrl=https://dblp.org/rec/conf/icteri/NemchenkoKP18
}}
==Neuro-Fuzzy Model of Development Forecasting and Effective Agrarian Sector
Transformations of Ukraine==
https://ceur-ws.org/Vol-2104/paper_201.pdf
Neuro-Fuzzy Model of Development Forecasting and
Effective Agrarian Sector Transformations of Ukraine
Diana Nemchenko1, Vitaliy Kobets1[0000-0002-4386-4103], and Larisa Potravka2
1
Kherson State University, 27, Universitetska st., Kherson, 73000 Ukraine
diana.alexandrovna03@gmail.com, vkobets@kse.org.ua
2
Kherson State Agricultural University, 23 Stretenskaya st., Kherson, 73000 Ukraine
potravka@rambler.ru
Abstract. Research goals and objective: to predict the economic dynamics of
the synergetic transformation model of Ukrainian agrarian sector using a neural
network on fuzzy data.
The object of research: Neuro-Fuzzy Model of Economic Forecasting.
The subject of research: forecasting the economic dynamics of the synergetic
transformation model of Ukrainian agrarian sector using a neural network on
fuzzy data.
Research Methods are neuro model, fuzzy logic, assessment of the risk of Vo-
ronov and Maksimov
Results of the research: We can say that the risk of this forecast, predicted by
the neural network, is "very low", we can definitely trust the forecast, and the
risk is calculated by the equation of the neuroregression "low", which indicates
that we can trust the forecast, but with caution and further monitoring.
Keywords: neuro model, fuzzy logic, economic forecasting.
1 Introduction
Application of optimization methods for fuzzy data is impossible, that`s why neuro-
fuzzy simulation is used as a mathematical methodology, which makes it possible to
put forward and solve even those problems which have no complete statistics or in
case there are only qualitative factors ensuring the possibility of adapting economic
and mathematical models to changing economic conditions.
The purpose of the paper is to predict the economic dynamics of the synergetic
transformation model of Ukrainian agrarian sector using a neural network on fuzzy
data. Determining the size of synergistic effect of economic, ecological and social
nature requires a mathematical interpretation with the use of up-to-date information
technology, since the calculation of synergy in economic system is complicated by the
random nature of economic phenomena in the conditions of transformation processes.
The development of scenarios for transforming the agrarian sector of the economy is
possible only with the use of information technology.
The paper is organized as follows: part 2 describes related works concerning neu-
ro-fuzzy models; part 3 describes Neuro-model "Nova Troya"; part 4 describes the
results of the neuromodulation; the last part concludes.
2 Related Works
In market conditions any economic agent during their activities inevitably faces un-
certainty. Even a professional is not able to predict changes that may occur in an un-
certain external environment. Simplification of economic system model in the frame-
work of traditional methods will inevitably lead to inadequacy of the resulting deci-
sions due to incomplete consideration of an uncertainty of internal and external sys-
tem environment. Consequently, the construction of accurate mathematical models of
innovative development of economic industries, fit for implementation in software
applications to solve analytical tasks of decision-making and its support, based on the
use of traditional methods, can either be difficult or impossible at all [1].
An alternative way of simulating the behavior of complex economic systems is the
assumption of their fuzziness when describing them. This statement is based on the
principle of incompatibility of accuracy and meaningfulness [2]. Thus, the approach
to solving economic problems of decision-making support has to be based on the fact
that the key elements are certain fuzzy sets rather than numbers, but. Failure to take
into account this factor in the creation of applied mathematical and software forecast-
ing largely determines the shortcomings of modern technologies and systems for mak-
ing economic decisions. Fuzzy logic as a set of theory basics, methods, algorithms,
procedures and software is based on the use of fuzzy knowledge and expert assess-
ments for solving a wide range of tasks [3].
This results in the fact that a number that has a specific meaning for an expert
ceases to have one value (which requires traditional mathematics), but can be ex-
pressed by a set of values with its own probability. In this case, the probability reflects
the impact and strength of possible active factors. The interpretation of fuzzy numbers
is determined on a case-by-case basis and depends on the physical nature of these
numbers, as well as on the factors that affect them. The fuzzy method allows to dra-
matically reduce the number of computations, which in turn leads to an increase in the
speed of fuzzy systems [4]. Fuzzy logic is based on the concept of fuzzy set as an
object with a function of belonging of an element to a set that can acquire any values
in the interval [0, 1], besides 0 or 1.
Artificial Neural Network is a mathematical toolkit that is a universal reproducer of
complex nonlinear functional dependencies, which is based on the principles of the
work of biological neural structures. This toolkit is used in data analysis, time series
forecasting, signal processing, pattern recognition, etc. [5]. The structure of the artifi-
cial neuron is graphically presented in fig. 1.
Fig. 1. Block structure of artificial neuron
An artificial neuron consists of an adder and a functional converter. The adder per-
forms calculation of weighted signals that arrive via interneuronal connections from
other neurons or external input signals. The functional converter transforms the output
of the adder by the activation function of the given type. Both natural and artificial
neurons can be trained depending on the activity of the processes that take place in
them. Also, as a result of training, the weight of the interneuronal connections also
changes, which also affects the behavior of the corresponding neuron.
Advantages and disadvantages of neural networks are demonstrated in table 1.
Table 1. Advantages and disadvantages of neural networks
Advantages of neural networks Disadvantages and limitations of neural
networks
- adaptability to environmental changes; - - effective forecasting requires a certain
training on examples; minimum number of observation (about 100
observations);
- parallel processing of information; - significant time expenditures to achieve a
satisfactory result;
- insensibility to errors; - only specialists can prepare reliable inter-
pretation of the results;
- ability to generalize gained knowledge - learning algorithm can fall into the "trap"
of the so-called local minimum error, and
the best solution will not be obtained;
- inability of traditional artificial neural
networks to "explain" how they solve the
problem.
One of the areas for application of neural networks is the agrarian industry. In the
research works [6] agrarian industry of Ukraine was proved to be an important reserve
for the growth of the national economy. Based on the experience of European Union
and South-east Asian countries it was determined that their economic growth is a
reault of deep transformations , oriented towards ensuring the achievement of re-
search and development in order to optimize the use of resources [7].
The experience of such models application indicates the possibility to predict the
probable consequences of macroeconomic and industry decisions in the context of
preserving existing relationships [8].
Modern relationships require new variables to describe them in the economic sys-
tem, which involves the expansion of characterization methods such as neural model-
ing and fuzzy logic. The high degree of probability of changes in economic systems is
formed under impact of external factors, which makes it impossible to clearly define
the goals of the updated system. In this case, the experience of traditional simulation
is not enough, so the transition to the neural model of the description of reality be-
comes relevant. In the transformation model of agrarian sector, the synergetic ap-
proach reflects the result of the joint interaction of economic, financial, social and
institutional factors (fig.2).
Fig. 2. Transformation model of agrarian sector
The synergetic effect determined by us is the result of the impact of external fac-
tors. For example, «Е+F=α» under impact of the external factor creates the effect "в",
which with the resulting index "s" (as a result of «F + C = ω») forms the effect "B".
Result of the political and economic component creates the necessary conditions for
the effective functioning of the agricultural sector of the economy [9].
Taking into account the results of the study, we believe that the synergetic effect of
transformations (S) has to be the sum of the synergetic effect of the components,
which function is close to the maximum under determined level of risk (r). The risks
are natural, climatic, political, demographic, space threats, informational, ethnic, reli-
gious, cultural, social, military conflict risks, terrorism, etc. [10, 11].
3 Neuro-model "Nova Troya" in the problem "Inflation-
production" inflation
The "Nova Troya" is a neural network model of "inflation-production", which uses a
sample of 32 quarterly data for 2009-2016, formed on the basis of the financial state-
ments of PLK TH "Nova Troya" (Ukraine, Kherson region Novotroitsk).
The task is to build a network based on the architecture of a multilayer perceptron
using the Excel Neural Package architecture, which bases our data and forms the link
between the indicators of economic growth of this enterprise (inputs of the model)
and the level of inflation (output of the model), estimated through the quarterly con-
sumer price index (CPI). The model will be used to predict the development of this
enterprise for future periods of time (quarters).
Stage One. Introducing the source and placing them in Excel.
Table 2. Output data
Stage Two. Using Neural Analysis, we describe the placement of table 3.
Fig. 3. Transformation model of agrarian sector
Stage Three. Identification of inputs (fig. 4)
Fig. 4. Identification of inputs
The All Data window displays a complete list of 8 parameters of the model. As the
input parameters we select the last 7. To do this, we perform the Select All command,
which will carry all parameters to the Inputs window. Then we return the "extra" CPI
parameter back to the All Data window using the adjustment button <. The result
obtained is shown below in fig.4.
Stage Four. We perform the preprocessing of input data using the Normalization
function. This step allows you to get rid of unnecessary computational problems due
to the alignment of the range of variables. We choose the Mean / Variance option, in
which the data becomes dimensionless by subtraction of the average value and divi-
sion by their dispersion. Now all inputs are comparable in order of magnitude.
Fig. 5. Align the ranges of variables
Stage Five. Next, using the Select Outputs function, we select the output parameter
- CPI and normalize it with Mean/Variance result:
Fig. 6. Determine the outputs
Stage Six. We determine the significance of the input parameters. We use the Box-
counting function, and the system by itself, using the Boxcounting algorithm, will
determine the statistical significance of the inputs for the specified outputs. In the
Boxcounting results window in graphical form we will see that the most significant
parameters are x2 and x6 (≈0.4 and 0.3, respectively), and the significance of the
parameters x1 and x5 is close to zero and they are insignificant in terms of the effect
on the resulting variable. . The rest of the variables occupy an intermediate position
by significance. The values of the normalization parameters are shown below: mean
predictability = 0.095 and variance (variance) = 0.109. The more their ratio is differ-
ent from 1, the better the predicted power of the model is. We calculate the ratio Av-
erage / Dispersion ≈ 0.87. Apparently, it is close to one, that is, the predicted strength
of this network is low. Reducing the number of inputs allows you to shorten the train-
ing time of the neural network or allows you to increase its nonlinear properties. So
let's go back to the main window and remove the insignificant entries x1 and x5 from
the list (fig.7).
Fig. 7. Exclusion of insignificant inputs
Stage Seven. Find the equation of neuroregression for CPI in following form:
117.81 1.7 ∙ 10 ∙ 3.003 ∙ 10 ∙ 1.76 ∙ 10 ∙
(1)
1.99 ∙ ,
where: - cost, - sales profit, - balance profit, - total profitability. For com-
parison purposes, we estimated the equation of multiple regression:
117.59 1.96 ∙ 10 ∙ 4.79 ∙ 10 ∙ 1.022 ∙ 10 ∙
(2)
1.96 ∙ .
The relevant statistics are shown in table 3.
Table 3. Relevant statistics of neuroregression
Neurostatistical indicator Meaning
Number of training examples, n 32
The number of independent variables, k 4
A common neuromuscular disorder, 1983.74
Σ(yn - ya )
Regression error, Σ(yс – ya)2 648.12
Regression error, 1335.62
Σe(t)2=Σ Σ(yn - ya )2- Σ(yс – ya)2
Determination coefficient, 0.33
R2=1-Σe(t) 2/ Σ(yn - ya )2)
Coefficient of multiple correlation, R= √R 2 0.57
Standard Error 7.03
Fisher test, F = (R 2/k)/((1- R 2 )(n-k-1)) 3.28
The significance of the Fisher test 0.03
The critical value of Fisher test with a confidence probability of 0.95, ν1 = k = 4,
ν2 = n-k-1 = 27 is 2.73. Since Ffact = 3.28 > Ftab = 2.73, the regression is adequate. To
estimate the independence of errors, we calculated the Durbin Watson criterion: d = Σ
(e (t) -e (t-1)) 2 / Σe (t) 2 = 1.41777E-05. As critical table levels for n = 32 and k = 4
for a significance level of 5%, we got critical values di = 1,18 and du = 1,73. The
calculated value did not fall into the interval, i.e. estimates can be considered inde-
pendent. The following table shows the calculation of the coefficients of the neuro-
regression equation and the Student's statistics (t-criterion):
Table 4. Coefficients of the neuroregression equation and Student's statistics
Coefficient
Variable of regres- Standard Error Student's test
sion
CPI (Output) 117.81 3.42 34.42
Cost price 1.70129E- 3.05E-06 0.5578
06
Profit from sales -3.00265E- 5.598E-05 -0.5363
05
Balance sheet 1.75885E- 5.599E-05 0.3141
profit 05
Profitability total -1.99 0.94 -2.12
Stage Eight. We create a simple two-layer neural network (with one hidden layer)
and architecture "4-3-1", using the Create Net command.
In the Network Constructor window, we define the network structure:
• number of layers without input (Number of layer) = 2;
• number of inputs (Number of inputs) = 4;
• the number of neurons in the 1st layer (Layer1, neuron) = 3;
• order of nonlinearity of the first layer (order) = 1;
• type of output function of the first layer (function) = tanh;
• the number of neurons in the second layer (neurons) = 1;
• the order of nonlinearity of the second layer (order) = 1;
• type of output function of the 2nd layer (function) = linear.
Fig. 8. Preparing for the creation of a neural network
We get the following neuron network:
Fig. 9. Received neural network
Stage nine: Preparation for training the neural network. Before training we set the
test set from the whole set of learning examples. Examples from this set will not par-
ticipate in the training. They will serve as a base for building the estimates for the
predicted properties of the trained network. With Edit test set in the window we set
the size of the test sample (Number of test examples) = 0, and its character set in the
random sample set (Random test set) (fig.10).
Fig. 10. Preparation for training the neural network
Stage ten: Training the neural network. In the course of the network learning you
can see the change in the parameters in the field of training information (Training
Info). We wait for the learning process to stop by itself or interrupt it artificially by
pressing Stop Training button (fig.11).
Fig. 11. Teaching the neural network
The learning process stopped by itself. As you can see, 9336 epoches have passed
at the moment, and the current training error is 0.2.
Learning outcomes can be assessed visually on the Network responses graph. The
corresponding window is shown below (green dotted line shows network feedback,
and orange one marks real data).
Fig. 12. Results of training of the neural network
Stage Eleven. Using Output all data, we output the results of the neural network
(table 3).
Table 3. Output data
As it can be seen from the table, the results of neuromodeling are well approximated
by actual data and the total square error is only 0.01%. The approximate regression
equation obtained on the basis of neuromodeling, of course, contains a big mistake,
but also only 0.15% (15 times worse). Graphic illustration of the table is given below.
Fig. 13. Comparison of the actual inflation rate with neuroforecast
4 Summing up the results of the neuromodulation using the
method of fuzzy logic
Let’s consider integral assessment of the risk of V & M (Voronov and Maksimov):
Rating:
• accepts values from 0 to 1;
• every investor, based on his/hers investment preferences, can classify the value by
allocating a segment of unacceptable risk values for themselves.
Advantages of the method:
• the full spectrum of possible scenarios for the investment process is formed on the
basis of the fuzzy sets theory;
• the decision is made on the basis of the whole set of assessments rather than two
assessments of the effectiveness of the project;
• the expected efficiency of the project is not a point indicator, but a field of interval
values with its distribution of expectations, characterized by the function of be-
longing to the corresponding fuzzy number.
Using the following graduation (table 5):
Table 5. Integral assessment of the risk of V & M
Gradation The degree of risk Decision regarding the
forecast
0 – 0,07 Very low Trust the forecast
0,07 – 0,15 Low Trust with caution and
further monimonitor-
ing
0,16 – 0,35 Average Trust with limitations
0,36 – 0,4 High Reject and view fore-
cast
> 0,40 Very high Do not trust the fore-
cast
0 – 0,07 Very low Trust the forecast
We can say that the risk of this forecast, predicted by the neural network, is "very
low", we can definitely trust the forecast, and the risk is calculated by the equation of
the neuroregression "low", which indicates that we can trust the forecast, but with
caution and further monitoring.
5 Conclusion
Using the Excel Neural Package, a developed network is based on the architecture of
the multilayer perceptron, which analyzes the data and forms the link between the
indicators of economic growth of analyzed enterprise.
The analysis of existing diagnostic technique of CPI (consumer price index) and
the assessment of the financial performance of the enterprise were carried out. It al-
lowed conducting a comprehensive financial and economic analysis of the enterprise
with using of fuzzy logic tools, which will enable the formation of an economic and
mathematical model taking into account the specificity of the enterprise.
Одним з найбільш ефективних математичних інструментів, спрямованих на
формалізацію і обробку невизначеної інформації, що інтегрує сучасні підходи і
методи, є теорія нечіткої логіки. Даний математичний апарат дозволяє розгля-
нути різні види невизначеності та отримати новий, якісно кращий прогноз ро-
звитку економічних систем.
One of the most effective mathematical tools assigned at formalizing and pro-
cessing of uncertain information which integrates modern approaches and methods is
fuzzy logic theory. This mathematical instrument allows us to consider various types
of uncertainty and get a new, qualitatively better forecast of the development of eco-
nomic systems.
References
1. Altunin, A.E., Semukhin, M.V.: Models and algorithms of decision-making under fuzzy
conditions. Tyumen State University, Tyumen (2000).
2. Aliev, R.A., Abdikeev, N.M., Shakhnazarov, M.M.: Production systems with artificial
intelligence. Radio and communication, Moscow (1990).
3. Nguyen, H.T., Walker, A.: First Course in Fuzzy Logic. CRC Press, USA (1999).
4. Pospelov, D.A.: Fuzzy sets in models of control and artificial intelligence. Science, Mos-
cow (1986).
5. Kobets V., Weissblut A.: (2018) Simulation Agent-Based Model of Heterogeneous Firms
Through Software Module. In: Bassiliades N. et al. (eds) Information and Communication
Technologies in Education, Research, and Industrial Applications. ICTERI 2017. Com-
munications in Computer and Information Science, Vol 826. Springer, Cham, pp. 230-254
doi.org/10.1007/978-3-319-76168-8_11
6. Potravka, L.O.: Estimation of synergetic effect of structural transformations of the agro-
food sector of Ukraine. Visnyk of Odessa National University named after Mechnikov 19
(20), 71--74 (2015).
7. Melnik, A., Vasina, A.: Structural transformation of the national economy of Ukraine as
factors of modernization of the institutional basis of its development. Journal of the Euro-
pean Economy. TNEU 9 (1), 40—64 (2010).
8. Moiseev, N.N.: The mathematician puts experiment. Nauka, Moscow (1979).
9. Potravka, L.O.: Synergy of transformations of the agrarian sector of Ukraine. KSAU,
Kherson (2016)
10. Kapitsa, S.P., Kurdyumov, S.P., Malinetsky, G.G.: Synergetics and forecasts of the future.
Nauka, Moscow (2003).
11. Kobets, V., Poltoratskiy, M.: Using an Evolutionary Algorithm to Improve Investment
Strategies for Industries in an Economic System (2016), CEUR Workshop Proceedings,
vol. 1614, P. 485-501 (Indexed by: Sci Verse Scopus, DBLP, Google Scholar). Available:
CEUR-WS.org/Vol-1614/ICTERI-2016-CEUR-WS-Volume.pdf