UNDERSTANDING THE THEORY, MODELS, AND APPLICATIONS

OF ARTIFICIAL NEURAL NETWORK

ABSTRACT

The primary objective of the Artificial Neural Network (ANN) is to build useful ‘computers' for serious
challenges and to reconstruct intelligent data methodological approaches like pattern recognition,
classification, and generalization through using simple, distributed, and robust processing units known
as artificial neurons. ANNs are parallel implementations of non-linear static-dynamic systems that are
fine-grained. The considerable degree of interconnectivity that provides neurons their great computing
capacity through their vast parallel-distributed structure gives ANNs their intelligence and ability to
tackle difficult issues. The recent spike in demand in ANN is primarily due to the fact that ANN
algorithms and architectures may be deployed in real-time applications using VLSI technology. The
scope of ANN applications has exploded in recent years, fueled by both theoretical and practical
accomplishments across a wide range of fields. The theory, models, and applications of artificial neural
networks are briefly discussed.

Keywords: Black Box Modeling, Neural Network models, Neural Network applications

I. INTRODUCTION

Contemplating public political behavior [2] will show the future landscape of a society in which anyone
can virtually perceive the future space and commence to programmer and create policy of his or her job
career, political, economic, military, and other sections affairs while being made aware regarding
obtainable characteristics, parties, and applicant’s perspectives. Because so many political theorists
struggle to predict public political behavior in these types of elections and are unable to do otherwise
precisely utilizing statistical tools and social and political analyzing methods [1], data mining methods
like artificial neural networks have been used to look for potential rules in these types of data by
collecting data and necessary documents from voters.

Artificial neural networks are a type of current mathematical-computational approach used for solving
unexpected dynamic issues in established behavioral systems over an amount of time. Artificial neural
networks can solve issues by evolving to detect patterns in data that other computational and statistical
methods have failed to address. [3] Artificial neural networks are becoming frequently utilized because
of their capacity to predict. Given that this characteristic, as well as participation predictors and election
results, as some of the most intriguing difficulties for political scientists, in such cases artificial neural
networks can be utilized to predict political issues, despite the fact that it would be less commonly
employed by researchers in Political Science. Political scientists and academics have already been
looking for large models to predict results of the election in emerging countries. [4]

A. Predict Using Neural Networks

The recent trend of utilizing artificial neural networks to predict outcomes has resulted in a significant
rise in research activity in the last decade. Artificial neural networks are effective tools for predicting
exchange rates, evaluating economic time - series data, and dealing with stock and stock market
difficulties. In terms of performance standards for future time sequences prediction, the findings and
functionality of artificial neural networks have proven that somehow this model performs better than
common techniques. Many sophisticated statistical approaches are being created and applied for the
anticipating process in social contexts in recent years. Nevertheless, there seem to be two major issues
with these approaches. Personal statistical problems, power, as well as some analyses for a single and
multi-dimensional time series are all included. Artificial neural networks were found to be effective in
removing statistical issues in multidimensional time series prediction.

Due to a lack of precise and satisfactory models in the economic area, it becomes difficult to predict
macro-economic difficulties. The time series model of black-box model, which partially addresses the
economy, is by far the most adequate assessment for economic analysis. High noise levels, short time
series, and nonlinear effects are used to resolve time series issues using the regression approach, which
has become one of the most prominent ways of solving economic forecasts in recent decades. Artificial
neural networks can also be used to resolve issues and predict macroeconomic concerns using the
supplied answer. The researchers were given results of the experiment to predict the criterion of

industrial production in the United States by choosing sophisticated parameters in artificial neural
networks, wherein the given results of their numerical simulations indicate that artificial neural networks
perform much better unlike commonly utilized time series and regression approaches.
The adaptable neural network model is able to identify and anticipate the explicit relation involving
democratization and universal collision theory in the political sphere, but the logical model does not.
Marchi et.al proposed the logic model. Researchers found it to outperform artificial neural networks,
although Nathaniel Beck et al. denied. They proceeded by analyzing Grynaviski, Gelpi, and Marchi's
ideas, emphasising that their points of view are perfectly compatible with theirs. Their shared opinion is
that major standards should be used to evaluate global research in order to predict out-of-sample
functionalities. They also said that Marchi et al's other results are erroneous. They agreed that Marchi et
al's use of the identical assessment criteria for each one of the two models was founded on logic model
supremacy above the other neural networks that offered challenges and rejected their hypothesis.

In the tourism industry, artificial neural network models outperformed other common forecasting
approaches such as time series and regression techniques in terms of anticipating international tourism
demand. It has been authorized in a study conducted by Rob Law and Norman Au. To predict a Japanese
travel to Hong Kong, the authors utilized an ordered Feed-Forward network, which is a type of neural
network. The input layer of a network consists of six service pricing nodes, average lodging rates,
currency rates, population, and internal gross expenditures, while the output layer of a node represents
travel needs from Japan to Hong Kong. The neural network model provides amazing results than most
other techniques, according to the experimental data collected from a research of the period interval
from 1967 to 1996.

In the field of environmental science, neural networks may be utilized to tackle difficulties and concerns
related to real-world process modelling by creating a composite architecture using meta-heuristic
algorithms. Non-linear multivariate space and chaos theory are two of these issues. Even though it is
hard to design neural networks for maximum efficiency, Harry Niska et al used a multi-layer perceptron
model to predict Nitrogen Dioxide during the day in a densely populated traffic station in Helsinki city
by combining genetic algorithms in tandem to choose input and design high-level architecture. It
demonstrated that mixed networks can solve the practical difficulties of constructing neural networks for
high efficiency [5].

In the realm of information technology, neural networks have proven that short-term intervals may
predict traffic rate in top-in networks superior than modified techniques of time series. Paulo Cortez et
al. conducted numerous tests to measure the reliability of the aforementioned techniques' predictive
performance using real-world data and associated information from internet service providers. The
studies were carried out at three time intervals of 5 minutes, 1 hour, and a day. However, 1-day
anticipation time series produced superior outcomes, and neural group tests have demonstrated that it
can produce the greatest results for short-term time series [6].

B. ANN Theory and Models

Different types of Neural Networks (NN) have also been developed, but they all share three elements in
common: the individual neuron, the architecture (connections between neurons), and the learning
method. Each category limits the types of connections that can be made. It might, for instance, state that
if one neuron is linked to another, the second neuron cannot connect to the first neuron again. The
structure or layout of the neural network refers to the many types of connections that may be made.

One or more layers of neurons make up a neural network. The output from the units in one layer is only
permitted to activate neurons in the next layer in a wide variety of NN models, including such
Perceptron, Linear Associator, Multi-layer feed-forward networks using Back-Propagation (BP)
learning, the Boltzmann machine, and the Grossberg model. In other models, such as Kohonen nets and
the Hopfield model, meanwhile, the signal can trigger neurons in the very same layer.

The numerical strength of a synaptic connection between two neurons is known as synaptic weight or
adaptation coefficient. Positive (excitory), negative (inhibitory), or zero (no connection) values can be
used to describe the strength of interconnectivity. The weight influences the shape of the signal that is
passed from one neuron to the next, therefore coding the network's knowledge. Whenever the
cumulative excitation surpasses the cumulative inhibition by a certain amount, termed the threshold (T),
usually 40 mV, the neuron fires, sending messages to neighbouring neurons. Only a few of the networks
offer real-time responses. Other networks require more opportunity to answer and are defined by their
time-domain behavior, which we call brain dynamics. The period of latent summation is the temporal
gap between when inputs are supplied and when neurons produce output.

If a neuron's threshold and input weights can be changed, it is considered to be "trainable." Neurons are
provided with inputs. If the neurons do not provide the expected output (as decided by us), then
something went wrong. To account for the discrepancy, various weights and thresholds must be
adjusted. The learning (or training) algorithm is the set of rules that control how well these changes will
take place. The method in which the modification of synaptic weights of a neuron is stated differs
amongst learning algorithms.
The network's weights are gradually changed in order to enhance a predetermined performance metric
over time. The cognitive development may best be described as a "search" for an answer in a multi-
dimensional weight space that gradually optimizes a pre-specified goal function. After each iteration of
the learning process, the NN has a better understanding of its surroundings.

C. ANN Applications

Since their inception, artificial neural networks (ANNs) have been used in a wide range of fields. The
general uses of several ANN models are listed in Table 1 in chronological order. The following are only
a handful of the many uses that have been described in the literature:

Alarm Processing

Engineers are required to swiftly assess numerous choices and adopt the best course of action in an
emergency. The quantity of real-time messages (alarms) received on the VDUs, on the other hand, is far
too great for the time available to evaluate them. A useful operational assistance has been discovered in
the processing of such alerts in real-time and notifying the operator to the root cause or the most
important of these alarms. For this type of alert processing, ANNs have been used. For this aim, Chan
[7] suggested a three-layer Perceptron network, which showed encouraging simulation results. In this
application, a trained ANN's quick reaction and generalization skills come in handy.

Eddy current analysis

The numerical solution of Integra-differential equations is required for the analysis of eddy current
losses. It is computationally costly to discretize these equations and solve them using finite-element
techniques. [8] Describe a cellular ANN that generated a technique for solving these equations that was
quicker, less computationally costly, and simpler. As an alternative to finite-element techniques, they
suggested a cellular NN.SPICE was used to mimic the cellular networks. In a time-varying magnetic
field, the cellular network computed eddy currents and eddy current losses in a source current carrying
conductor. This approach allows for a wide range of structural analysis, electromagnetic field
calculations, and other applications.

Harmonic source monitoring

[9, 10] describe the detection and monitoring of harmonic sources in non-linear load systems. This
method presupposes that the system's harmonics can be measured directly. Multiple three-layer
Perceptrons were used by these researchers. Simulation data for various load circumstances were used to
train the ANNs. To locate and monitor the source of the harmonics, the ANNs were utilized in
combination with a state estimator. This method proved successful in identifying a previously
unidentified harmonic source.

Applications in nuclear power plants

Researchers at the University of Tennessee explored the possible uses of ANNs in increasing the
effectiveness and productivity of nuclear power reactors as part of a study funded by the US Department
of Energy [11, 12, and 13]. The following topics were investigated: assessment of particular faulty
situations, recognition of a change in mode of operation, signal validation, check valve monitoring,
modelling of plant thermodynamics, plant parameter monitoring, plant vibration analysis, and so on.

II. OBJECTIVE OF THE STUDY

The paper's main objective is to investigate neural networks and their applications.

III. LITERATURE REVIEW

The current Special Issue, titled "Advances in Neural Networks Research: IJCNN2009," presents a
current state-of-the-art assessment of neural networks. It contains 39 papers from the 2009 International
Joint Conference on Neural Networks in chosen topics (IJCNN2009). The International Joint
Conference on Neural Networks and Computational Intelligence (IJCNN2009) was held in Atlanta,
Georgia, USA, from June 14 to 19, 2009, and it was an excellent example of collaborative efforts
between the International Neural Networks Society and the IEEE Computational Intelligence
Society.Computational intelligence and machine learning, hybrid methods, nonlinear dynamics and
chaos, different soft communications technology, intelligent signal processing and pattern recognition,
bio-informatics and bio-medicine, and engineering applications are among the topics covered in this
issue. [14]

Artificial neural networks have been shown to be extremely successful in predicting and analyzing flows
that traditional techniques and statics are unable to address. In this paper, we show so using a two-layer
feed forward network with such a tan-sigmoid transmission function in the input and output layers, we
can predict the public participation rate in future presidential elections in the Islamic Republic of Iran
with 91 percent respectively in Kohgiloye and Boyerahmad Provinces. Our assertions have been
validated using participation evaluation criteria including confusion matrix and ROC diagrams.

In this study, [15] we proposed a technique that uses Artificial Neural Networks and expert systems to
understand deeply for the learner model in the Linear Programming Intelligent Tutoring System (LP-
ITS) for determining the learners' academic performance level so that the appropriate level of difficulty
of linear programming problems can be offered to them to solve. To predict academic achievement, LP-
ITS utilises a Feed forward Back-propagation algorithm that has been trained with data from a set of
students. Additionally, LP-ITS employs an Expert System to determine the appropriate degree of
difficulty for the learner's expected academic achievement. Numerous tests were performed to see if
real-time data was adhered to. The reliability of forecasting the learners' achievement is quite good,
indicating that the Artificial Neural Network is capable of making appropriate predictions.

IV. CONCLUSION

Even when the training data contains flaws, the ANN can generate a generalized solution to the issue
that is not the same as the one used for training and give legitimate results. These features contribute to
making NN an effective tool for modelling situations involving functional connections that are unclear
or likely to change over time.

Another instance where the NN technique might be beneficial is if the time necessary to create a
solution is essential, like real-time applications that require a large number of solutions in a short period
of time. Even when other approaches that can generate more optimum results are available, the capacity
of a NN to produce rapid solutions regardless of the complexity of the problem makes them desirable.
[2]
ANN do not offer a formal representation of the relationship between input and output data, in contrast
to classical methods or expert systems, which seek to codify knowledge and build partial qualitative
system models. As a result, stastical tests must be used to evaluate the ANN's performance.

REFERENCES
[1] Sangar, A. B., Khaze, S. R., &Ebrahimi, L (2013), Participation Anticipating In Elections Using
Data Mining Methods , International Journal On Cybernetics & Informatics,2(2),47-
60.International Journal of Information Technology, Modeling and Computing (IJITMC) Vol.1,
No.3,August 2013 31
[2] Gharehchopogh, F. S. &Khaze, S. R. (2012). Data Mining Application for Cyber Space Users
Tendency In Blog Writing: A Case Study, International journal of computer applications, 47(18),
40- 46.
[3] Gill, G. S. (2008), Election Result Forecasting Using Two Layer Perceptron Network ,Journal of
Theoretical and Applied Information Technology, 47(11),1019-1024
[4] Caleiro, A. (2005), How to Classify a Government? Can a Neural Network do it? , University of
Evora, Economics Working Papers.
[5] [10] Niska, H., Hiltunen, T., Karppinen, A., Ruuskanen, J., &Kolehmainen, M. (2004), Evolving
The Neural Network Model For Forecasting Air Pollution Time Series, Engineering
Applications of Artificial Intelligence, 17(2), 159-167.
[6] Cortez, P., Rio, M., Rocha, M., & Sousa, P. (2012), Multi‐Scale Internet Traffic Forecasting
Using Neural Networks And Time Series Methods, Expert Systems, 29(2), 143-155.
[7] E. H. Chan, “Applications of neural network computing in intelligent arm processing,” Proc.
IEEE PICA Conf., pp. 246-251, Seattle, WA, USA, 1989.
[8] Y. J. Feria, J.D. McPherson and D.J. Rolling, “Cellular neural networks for eddy current
problems,” IEEE Trans. Power Delivery, vol 6, pp. 187-191, 1991.
[9] R. K. Hartana and G. G. Richards, “Constrained neural network based identification of
harmonic sources,” Conf. IEEE IAS annual meeting, pp. 1743- 1748, 1990.
[10] R. K. Hartana and G. G. Richards, “Harmonic source monitoring and identification using
neural networks,” IEEE Trans. on power systems, vol 5, pp. 1098-1104, May 2003.
[11] R. E. Uhrig, “Potential applications of neural networks to nuclear power plants,” Proc. Am.
Power conf., vol 53, pp.946-951, 1991.
[12] Y. Izumi and H. Taoka, “Neural networks and its applications to power system,” Electr. Eng.
In jpn., vol 109, pp. 230, 1989.
[13] A. Mathur and T. Samad, “Neural networks-a tutorial for the power industry,” Proc. Am.
Conf., vol 52, pp. 239-244, 1990.

[14] Kozma, R., Bressler, S., Perlovsky, L., & Kumar Venayagamoorthy, G. (2009). Neural
Networks. ELSEVIER, 22(489-490). https://www.journals.elsevier.com/neural-networks
[15] Abu Naser, S. S. (2012). Predicting Learners Performance Using Artificial Neural Networks
in Linear Programming Intelligent Tutoring System. International Journal of Artificial
Intelligence & Applications, 3(2), 65–73. https://doi.org/10.5121/ijaia.2012.3206