A

MAJOR PROJECT REPORT

ON

MACHINE LEARNING TECHNIQUES FOR OPTIMIZING

DESIGN OF DOUBLE T- SHAPED MONOPOLE ANTENNA
Submitted in partial fulfillment of the requirement for the award of degree of

BACHELOR OF TECHNOLOGY

in

ELECTRONICS AND COMMUNICATION ENGINEERING

By

A.SOWMYA 17QK1A0402

Under the supervisor of

M.SRAVANTHI(M.TECH)
Assistant Professor

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

JYOTHISHMATHI INSTITUTE OF TECHNOLOGICAL SCIENCES

(Recognized by AICTE & Affiliated to JNTU, Hyderabad)

Ramakrishna Colony, Karimnagar -505481

2017-2021
 JYOTHISHMATHI INSTITUTE OF TECHNOLOGICAL SCIENCES

Ramakrishna Colony, Karimnagar-505481

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

CERTIFICATE
This is to certify that this project work entitled “Machine Learning
Techniques for Optimizing Design of Double T-shaped Monopole
Antenna” is a bonafide workcarried out by A.SOWMYA (17QK1A0402)
in partial fulfillment of the requirements for the award of Bachelor of
Technology in Electronics and Communication Engineering during the
period of 2017-2021 underthe guidance and supervision.

SUPERVISOR HEAD OF THE DEPARTMENT

M.SRAVANTHI (M.TECH) M. RAMA KRISHNA,M.TECH.,
Assistant Professor Associate Professor
Department of Electronics and Department of Electronics and
Communication engineering Communication engineering
JITS, Karimnagar. JITS, Karimnagar.

Submitted for the Project Viva Voce Examination held on

 DECLARATION
I, A.SOWMYA (17QK1A0402) hereby declare that the project entitled
“Machine Learning Techniques for Optimizing Design of Double T-
shaped Monopole Antenna”, submitted in the partial fulfillment of the
requirements for the award of Bachelor of Technology in Electronics and
Communication Engineering from Jyothishmathi Institute of
Technological Sciences, Ramakrishna Colony, Karimnagar,
recognized by AICTE & Affiliatedto JNTUH, Hyderabad is an authentic
work and has not been submitted toany other university or institute for
award of the Under Graduation.

A.SOWMYA
17QK1A0402
 ACKNOWLEDGEMENT

At first, I would like to express my most sincere gratitude to my

supervisor, M.SRAVANTHI(M.TECH) Assistant Professor Department
of ECE, Jyothishmathi Institute of Technological Sciences, Karimnagar
for his guidance, support and encouragement. His vast experience and deep
understanding of the subject proved to be immense help to me, and also his
profound viewpoints and extraordinary motivation enlightened me in
many ways. I just hope my thinking and working attitudes have been
shaped according to such outstanding qualities.
I wish to express my sincere thanks to M. RAMA KRISHNA,
M.TECH Associate Professor & Head , Jyothishmathi Institute of
Technological Sciences, Karimnagar for giving his valuable suggestions
and being an eminent guide for the completion of this project.
I wish to express my sincere thanks to Dr. P. K. VAISHALI, M.Tech.,
Ph.D, principal, Jyothishmathi Institute of Technological Sciences,

Karimnagar for providing the college facilities for the completion of the
project.
Finally, I thank to all the faculty members, management and supporting
staff of ECE department, and friends for their kind co- operation and
valuable help for completing the project. My special thanks to my father
Mr. A. KAMALAKAR RAO and my mother Mrs.A.SARITHA for their
love,understanding and encouragement throughout my Project. Finally, I
thank Almighty God for always being there to guide me through thick and
thin.

A.SOWMYA
17QK1A0402
 ABSTRACT

In this paper we propose using modern machine learning techniques

including least absolute shrinkage and selection operator (lasso), artificial neural
networks (ANN) and k-nearest neighbor methods (KNN), for antenna design
optimization. The automated techniques are shown to provide an efficient,
flexible, and reliable framework to identify optimal design parameters for a
reference dual-band double T-shaped monopole antenna to achieve favorite
performance in terms of its two bands i.e. between 2.4-3.0 GHz and 5.15-5.6
GHz, respectively. In this paper, we also present a thorough study and
comparative analysis of the results predicted by these machine learning
techniques, with results obtained from HFSS to verify the accuracy of these
techniques.

Index Terms—Antenna Optimization, Experiment design, Linear regression,

lasso shrinkage, Machine learning, Optimization

i
 TABLE OF CONTENTS
CHAPTER NO. TITLE PAGE NO.
ABSTRACT I
LIST OF FIGURES IV
LIST OF TABLES V
LIST OF ABBREVIATIONS VI
1 INTRODUCTION 1
1.1 Machine Learning 3
1.1.1 History 5
1.1.2 Machine Learning techniques 7
1.1.3 Applications 13
1.1.4 Advantages and Disadvantages 15
1.1.5 Future Prosperity of Machine Learning 18
2 LITERATURE REVIEW 19
ANSYS HFSS 23
Ansys HFSS for antenna design 24
Applications 25
3 METHDOLOGY 27
3.1 Monopole antenna 27
Advantages of Monopole antenna 29
3.1.1
3.1.2 Drawbacks of Monopole antenna 30
3.1.3 Applications 31
3.1.4 Proposed antenna design parameters 32

4 RESULT AND CONCLUSION 20

4.1 Results 47
4.2 Conclusion 53

ii
 54
5
REFERENCES

iii
 LIST OF FIGURES

FIGURE NO. FIGURE NAME PAGE NO.

1.1.1 Lasso geometry 8
1.1.2 Artificial neurons networks 10
1.1.3 Example of knn classification 11
3.1.1 Basic monopole antenna 28
3.1.2 Proposed antenna design 33
3.1.3 ANN Architecture based on MLP 36
4.1 (a) FOM values with respect to change in l21 47
(b) FOM values with respect to change in l22 48
(c) FOM values with respect to change in w1 49
(d) FOM values with respect to change in w2 50
(e) FOM values with respect to change in w 51
(f) Simulated S11 Frequency 52

iv
 LIST OF TABLES

TABLE NO. TABLE NAME PAGE NO.

5.1 Predicted fom design parameter 41
values and computation line
Comparison
5.1 Sensitive analysis of predicted lasso 43
model

v
 LIST OF ABBREVIATIONS

ISM Industrial, Scientific and Medical

NFC Near Field Communication
FCC Federal Communications Commission
Wi-Fi Wireless Fidelity
RFID Radio Frequency Identification
VSWR Voltage Standing Wave Ratio
UWB Ultra Wide Band
WI-MAX Worldwide Interoperability for Microwave Access
SCADA Supervisory Control and Data Acquisition
LOS Line of Sight
NLOS Non-Line-Of-Sight
WMTS Wireless Medical Telemetry Service
ITU International Telecommunication Union
CAD Computer Aided Design
WLAN Wireless Local Area Network
IE3D Integral Equation for 3-Dimensional
IFA Inverted-F Antenna
OSA Open-Slot Antenna
DTH Direct To Home
MIC Microwave Integrated ircuits
ECM Electronic Counter Measures
ECCM Electronic Counter-Counter Measures
GPS Global Positioning System
ELMS Earth Limb Measurement Satellite
ISEE International Sun Earth Explorer

vi
SIR Shuttle Imaging Radar
SME Solar Mesospheric Explorer
ATMS Antenna Test and Measurement Society

vii
 CHAPTER 1

INTRODUCTION

The upcoming era of internet of things, (IoT) has enabled an immense growth in the
demand of application-specific antennas, which are needed for almost all electronic
devices. Hence, the requirement of a smart and efficient way of an-tenna designing
has become inevitable. Current antenna design relies heavily on the designers
empirical experiences and EM simulations. Traditional methods are inherently
inefficient and computationally intensive, making them impractical when there are
a large number of antenna design parameters to be optimized such as for 3-D printed
antennas [1]. To address challenges for designing complex 3-D structures, machine
learning (ML) techniques may be highly beneficial. ML has been widely used as an
indispensable data analysis and decision-making tool in a broad range of
applications ranging from hand-written digit recognition [2] to human genomics [3].
Researchers have also explored optimization of antenna structures by applying
heuristic optimization techniques like genetic algorithms, particle swarm
optimization to antenna designs [4]–[6] as well, but these algorithms search for the
optimal solution by analyzing the output on individual data points and generating
new and possibly better search directions until a global maxima or minima is
identified. On the other hand, machine learning (ML) refers to all techniques and
Optimization algorithms of analyzing the data and finding the hidden mathematical
relation in data such that we can relate the input behavior to the output behavior and
make future predictions or decisions using this relationship. The main advantage of
using ML techniques is once we have the relational model, we can predict the output
for any data point rather than aiming for global optimal and minima points only. This
property is very beneficial when we want to use the same data set for multiple
different goals. As described in [7], there is some early work on applying machine

1
learning techniques for antenna analysis and synthesis [8]–[11]. In [8], the
performance of support vector machines (SVMs) is investigated for designing of a
rectangular patch antenna and a rectangular patch array while in [9] methodology to
use SVMs for linear and nonlinear beamforming and parameter design for array and
electromagnetic applications is mentioned. Another ML technique, Artificial Neural
Network has also been applied in this field [10]-[11]. In [12], clustering method is
used to find the optimum position for shorting posts in microstrip patch array design
so as to achieve acceptable bandwidth, scan angle and polarization. As observed
from here, some studies for using ML techniques for antenna design optimization
have been conducted but a detailed analysis and systematic comparison of various
ML techniques for antennas have not been reported. The main contribution of this
work is to fill the gap by presenting new classes of ML-based methods for automated
antenna design optimization, evaluating their performance in terms of prediction
accuracy and robustness, and making comparisons with EM simulations. Our
discovery suggests that ML is a promising choice to provide automated,
computational feasible and practically effective approaches for antenna design. The
ultimate goal of this study is to further extend the proposed ideas to more complex
design of antennas and develop scalable and efficient algorithms to tackle
computational challenges, by handling a large number of design parameters. In this
paper, we propose using machine learning techniques for antenna design
optimization, and particularly consider artificial neural networks (ANNs), least
absolute shrinkage and selection operator (lasso), and k-nearest neighbor (kNN).
The feasibility of these new approaches for antenna design is demonstrated by their
applications to optimize a reference double T-shaped monopole antenna [13] as
shown in Fig. 1.

2
The initial work related to this was reported in [14] where only
the lasso technique was employed and targeted optimization function was weighted
sum of fractional bandwidths of two bands of the reference antenna. In this paper,
we use finer optimization function and explore more machine learning techniques
to optimize the performance of the reference antenna. This helps in achieving
better results as compared to [13] and [14]

3
1.1 MACHINE LEARNING

Machine learning (ML) is the study of computer algorithms that improve

automatically through experience and by the use of data.[1] It is seen as a part of
artificial intelligence. Machine learning algorithms build a model based on sample
data, known as "training data", in order to make predictions or decisions without
being explicitly programmed to do so.[2] Machine learning algorithms are used in a
wide variety of applications, such as in medicine, email filtering, and computer
vision, where it is difficult or unfeasible to develop conventional algorithms to
perform the needed tasks.[3]

A subset of machine learning is closely related to computational statistics, which

focuses on making predictions using computers; but not all machine learning is
statistical learning. The study of mathematical optimization delivers methods, theory
and application domains to the field of machine learning. Data mining is a related
field of study, focusing on exploratory data analysis through unsupervised
learning.[5][6] In its application across business problems, machine learning is also
referred to as predictive analytics.

Machine learning involves computers discovering how they can perform tasks
without being explicitly programmed to do so. It involves computers learning from
data provided so that they carry out certain tasks. For simple tasks assigned to
computers, it is possible to program algorithms telling the machine how to execute
all steps required to solve the problem at hand; on the computer's part, no learning
is needed. For more advanced tasks, it can be challenging for a human to manually
create the needed algorithms. In practice, it can turn out to be more effective to help

4
the machine develop its own algorithm, rather than having human programmers
specify every needed step.[7]

The discipline of machine learning employs various approaches to teach computers

to accomplish tasks where no fully satisfactory algorithm is available. In cases where
vast numbers of potential answers exist, one approach is to label some of the correct
answers as valid. This can then be used as training data for the computer to improve
the algorithm(s) it uses to determine correct answers. For example, to train a system
for the task of digital character recognition, the MNIST dataset of handwritten digits
has often been used.[7]

1.1.1 HISTORY
The term machine learning was coined in 1959 by Arthur Samuel, an American
IBMer and pioneer in the field of computer gaming and artificial intelligence.[8][9]
A representative book of the machine learning research during the 1960s was the
Nilsson's book on Learning Machines, dealing mostly with machine learning for
pattern classification.[10] Interest related to pattern recognition continued into the
1970s, as described by Duda and Hart in 1973.[11] In 1981 a report was given on
using teaching strategies so that a neural network learns to recognize 40 characters
(26 letters, 10 digits, and 4 special symbols) from a computer terminal.[12]

Tom M. Mitchell provided a widely quoted, more formal definition of the algorithms
studied in the machine learning field: "A computer program is said to learn from
experience E with respect to some class of tasks T and performance measure P if its
performance at tasks in T, as measured by P, improves with experience E."[13] This

5
definition of the tasks in which machine learning is concerned offers a fundamentally
operational definition rather than defining the field in cognitive terms. This follows
Alan Turing's proposal in his paper "Computing Machinery and Intelligence", in
which the question "Can machines think?" is replaced with the question "Can
machines do what we (as thinking entities) can do?".[14]

Modern day machine learning has two objectives, one is to classify data based on
models which have been developed, the other purpose is to make predictions for
future outcomes based on these models. A hypothetical algorithm specific to
classifying data may use computer vision of moles coupled with supervised learning
in order to train it to classify the cancerous moles. Where as, a machine learning
algorithm for stock trading may inform the trader of future potential predictions.[15]

Lasso was developed independently in geophysics literature in 1986, based on prior

work that used the penalty for both fitting and penalization of the coefficients.
Statistician Robert Tibshirani independently rediscovered and popularized it in
1996, based on Breiman's nonnegative garrote.[1] [4]

Prior to lasso, the most widely used method for choosing covariates was stepwise
selection. That approach only improves prediction accuracy in certain cases, such as
when only a few covariates have a strong relationship with the outcome. However,
in other cases, it can increase prediction error.

At the time, ridge regression was the most popular technique for improving
prediction accuracy. Ridge regression improves prediction error by shrinking the
sum of the squares of the regression coefficients to be less than a fixed value in order
to reduce over fitting, but it does not perform covariate selection and therefore does
6
not help to make the model more interpretable.

Lasso achieves both of these goals by forcing the sum of the absolute value of the
regression coefficients to be less than a fixed value, which forces certain
coefficients to zero, excluding them from impacting prediction. This idea is similar
to ridge regression, which also shrinks the size of the coefficients, however Ridge
Regression tends to set far fewer coefficients to zero.

The history of Artificial Intelligence (AI) began in antiquity, with myths, stories and
rumors of artificial beings endowed with intelligence or consciousness by master
craftsmen. The seeds of modern AI were planted by classical philosophers who
attempted to describe the process of human thinking as the mechanical manipulation
of symbols. This work culminated in the invention of the programmable digital
computer in the 1940s, a machine based on the abstract essence of mathematical
reasoning. This device and the ideas behind it inspired a handful of scientists to begin
seriously discussing the possibility of building an electronic brain.

1.1.2 MACHINE LEARNING TECHNIQUES

➢ LASSO (least absolute shrinkage and selection operator)

In statistics and machine learning, lasso (least absolute shrinkage and

selection operator; also Lasso or LASSO) is a regression analysis method
that performs both variable selection and regularization in order to
enhance the prediction accuracy and interpretability of the resulting
statistical model. It was originally introduced in geophysics,[1] and later
by Robert Tibshirani,[2] who coined the term. Lasso was originally
formulated for linear regression models. This simple case reveals a

7
substantial amount about the estimator. These include its relationship to
ridge regression and best subset selection and the connections between
lasso coefficient estimates and so-called soft thresholding. It also reveals
that (like standard linear regression) the coefficient estimates do not need
to be unique if covariates are collinear.

Though originally defined for linear regression, lasso regularization is

easily extended to other statistical models including generalized linear
models, generalized estimating equations, proportional hazards models,
and M-estimators.[2][3] Lasso's ability to perform subset selection relies
on the form of the constraint and has a variety of interpretations including
in terms of geometry, Bayesian statistics and convex analysis.

The LASSO is closely related to basis pursuit denoisin

8
 Fig1.1.1 Lasso geometry

LASSO Geometry

LASSO forms a diamond shape in the plot for its constraint region, as shown in the
image below. The diamond shape includes corners, unlike the circular shape formed
by ridge regression. The proximity of the first point to the corner shows that the
model comes with one coefficient, which is equal to zero.

The ridge regression constraints region forms a circular shape that includes no
corners similar to that formed by the LASSO constraints region when plotted. The
ridge regression coefficients can, therefore, not be equal to zero.

➢ Artificial neural networks (ANNs)

Artificial neural networks (ANNs), usually simply called neural networks (NNs), are
computing systems vaguely inspired by the biological neural networks that
constitute animal brains.

9
 Fig 1.1.2 An artificial neural network is an interconnected group of nodes,
Inspired by a simplification of neurons in a brain. Here, each circular
node represents an artificial neuron and an arrow represents a connection
from the output of one artificial neuron to the input of another.

An artificial neural network is an interconnected group of nodes, inspired by a

simplification of neurons in a brain. Here, each circular node represents an artificial
neuron and an arrow represents a connection from the output of one artificial neuron
to the input of another.

An ANN is based on a collection of connected units or nodes called artificial

neurons, which loosely model the neurons in a biological brain. Each connection,
like the synapses in a biological brain, can transmit a signal to other neurons. An
artificial neuron that receives a signal then processes it and can signal neurons
connected to it. The "signal" at a connection is a real number, and the output of each
neuron is computed by some non-linear function of the sum of its inputs. The

10
connections are called edges. Neurons and edges typically have a weight that adjusts
as learning proceeds. The weight increases or decreases the strength of the signal at
a connection. Neurons may have a threshold such that a signal is sent only if the
aggregate signal crosses that threshold. Typically, neurons are aggregated into
layers. Different layers may perform different transformations on their inputs.
Signals travel from the first layer (the input layer), to the last layer (the output layer),
possibly after traversing the layers multiple times.

➢ k-nearest neighbors algorithm (k-NN)

In statistics, the k-nearest neighbors algorithm (k-NN) is a non-parametric

classification method first developed by Evelyn Fix and Joseph Hodges in 1951,[1]
and later expanded by Thomas Cover.[2] It is used for classification and regression.
In both cases, the input consists of the k closest training examples in data set. The
output depends on whether k-NN is used for classification or regression:

11
Fig 1.1.3 Example of k-NN classification .The test sample (green dot) should be classified either to blue
squares or to red triangle. If k=3 (solid line circle) it is assigned to the red triangle because there are 2
triangle and only 1 square inside the inner circle. If k=5 (dashed line circle) it is assigned to the blue
squares(3squares vs.2 triangle inside the outer circle).

In k-NN classification, the output is a class membership. An object is classified by

a plurality vote of its neighbors, with the object being assigned to the class most
common among its k nearest neighbors (k is a positive integer, typically small). If
k = 1, then the object is simply assigned to the class of that single nearest neighbor.
In k-NN regression, the output is the property value for the object. This value is the
average of the values of k nearest neighbors.
k-NN is a type of classification where the function is only approximated locally and
all computation is deferred until function evaluation. Since this algorithm relies on
distance for classification, if the features represent different physical units or come
in vastly different scales then normalizing the training data can improve its accuracy
dramatically.[3][4]

Both for classification and regression, a useful technique can be to assign weights to
the contributions of the neighbors, so that the nearer neighbors contribute more to
the average than the more distant ones. For example, a common weighting scheme
consists in giving each neighbor a weight of 1/d, where d is the distance to the
neighbor.[5]

The neighbors are taken from a set of objects for which the class (for k-NN
classification) or the object property value (for k-NN regression) is known. This can
be thought of as the training set for the algorithm, though no explicit training step is
12
required.

A peculiarity of the k-NN algorithm is that it is sensitive to the local structure ofdata.
1.1.3 APPLICATIONS

➢ SPEECH RECOGNITION

Speech recognition is the translation of spoken words into the text. It is also known
as computer speech recognition or automatic speech recognition. Here, a software
application can recognize the words spoken in an audio clip or file, and then
subsequently convert the audio into a text file. The measurement in this application
can be a set of numbers that represent the speech signal. We can also segment the
speech signal by intensities in different time-frequency bands.

Speech recognition is used in the applications like voice user interface, voice
searches and more. Voice user interfaces include voice dialing, call routing, and
appliance control. It can also be used a simple data entry and the preparation of
structured documents.

➢ Facial Recognition Becomes A Reality

In 2006, the Face Recognition Grand Challenge – a National Institute of Standards

and Technology program – evaluated the popular face recognition algorithms of the
time. 3D face scans, iris images, and high-resolution face images were tested. Their
findings suggested the new algorithms were ten times more accurate than the facial

13
recognition algorithms from 2002 and 100 times more accurate than those from
1995. Some of the algorithms were able to outperform human participants in
recognizing faces and could uniquely identify identical twins.

In 2012, Google’s X Lab developed an ML algorithm that can autonomously browse

and find videos containing cats. In 2014, Facebook developed DeepFace, an
algorithm capable of recognizing or verifying individuals in photographs with the
same accuracy as humans.

➢ Medical Diagnosis

Machine learning can be used in the techniques and tools that can help in the
diagnosis of diseases. It is used for the analysis of the clinical parameters and their
combination for the prognosis example prediction of disease progression for the
extraction of medical knowledge for the outcome research, for therapy planning and
patient monitoring. These are the successful implementations of the machine
learning methods. It can help in the integration of computer-based systems in the
healthcare sector.

➢ Financial Services
Machine learning has a lot of potential in the financial and banking sector. It is the
driving force behind the popularity of the financial services. Machine learning can
help the banks, financial institutions to make smarter decisions. Machine learning
can help the financial services to spot an account closure before it occurs. It can also
track the spending pattern of the customers. Machine learning can also perform the
market analysis. Smart machines can be trained to track the spending patterns. The
14
algorithms can identify the tends easily and can react in real time.

➢ Prediction Systems
Machine learning can also be used in the prediction systems. Considering the loan
example, to compute the probability of a fault, the system will need to classify the
available data in groups. It is defined by a set of rules prescribed by the analysts.
Once the classification is done, we can calculate the probability of the fault. These
computations can compute across all the sectors for varied purposes. Making
predictions is one of the best machine learning applications.

➢ Regression
We can also implement machine learning in the regression as well. In regression, we
can use the principle of machine learning to optimize the parameters. It can also be
used to decrease the approximation error and calculate the closest possible outcome.
We can also use the machine learning for the function optimization. We can also
choose to alter the inputs in order to get the closest possible outcome.

1.1.4 ADVANTAGES

➢ Easily identifies trends and patterns

Machine Learning can review large volumes of data and discover specific trends and
patterns that would not be apparent to humans. For instance, for an e-commerce
website like Amazon, it serves to understand the browsing behaviors and purchase

15
histories of its users to help cater to the right products, deals, and reminders relevant
to them. It uses the results to reveal relevant advertisements to them.

➢ No human intervention needed (automation)

With ML, you don’t need to babysit your project every step of the way. Since it
means giving machines the ability to learn, it lets them make predictions and also
improve the algorithms on their own. A common example of this is anti-virus
softwares; they learn to filter new threats as they are recognized. ML is also good at
recognizing spam.

➢ Continuous Improvement
As ML algorithms gain experience, they keep improving in accuracy and efficiency.
This lets them make better decisions. Say you need to make a weather forecast
model. As the amount of data you have keeps growing, your algorithms learn to
make more accurate predictions faster.

1.1.4 DISADVANTAGES

➢ Data Acquisition
Machine Learning requires massive data sets to train on, and these should be
inclusive/unbiased, and of good quality. There can also be times where they must

16
wait for new data to be generated.

➢ Time and Resources

ML needs enough time to let the algorithms learn and develop enough to fulfill their
purpose with a considerable amount of accuracy and relevancy. It also needs massive
resources to function. This can mean additional requirements of computer power .

➢ High error-susceptibility
Machine Learning is autonomous but highly susceptible to errors. Suppose you train
an algorithm with data sets small enough to not be inclusive. You end up with biased
predictions coming from a biased training set. This leads to irrelevant advertisements
being displayed to customers. In the case of ML, such blunders can set off a chain
of errors that can go undetected for long periods of time. And when they do get
noticed, it takes quite some time to recognize the source of the issue, and even longer
to correct it.

➢ Interpretation
When the algorithms help in all these processes and give a resulting output. This
given output must be checked for any errors and the correction operation should be
followed to get the desired accuracy. And during the selection of this algorithm, we
must select that algorithm which you require for the purpose.

17
1.1.5 FUTURE PROSPERITY OF MACHINE LEARNING

While machine learning algorithms have been around for decades, they've attained
new popularity as artificial intelligence has grown in prominence. Deep learning
models, in particular, power today's most advanced AI applications.

Machine learning platforms are among enterprise technology's most competitive

realms, with most major vendors, including Amazon, Google, Microsoft, IBM and
others, racing to sign customers up for platform services that cover the spectrum of
machine learning activities, including data collection, data preparation, data
classification, model building, training and application deployment.

As machine learning continues to increase in importance to business operations and

AI becomes more practical in enterprise settings, the machine learning platform wars
will only intensify.

Continued research into deep learning and AI is increasingly focused on developing

more general applications. Today's AI models require extensive training in order to
produce an algorithm that is highly optimized to perform one task. But some
researchers are exploring ways to make models more flexible and are seeking
techniques that allow a machine to apply context learned from one task to future,
different tasks.

18
 CHAPTER 2

LITERATURE REVIEW

• H. Xin and M. Liang, “3-D printed microwave and THz devices using
polymer jetting techniques,” Proc. IEEE, vol. 105, no. 4, pp. 737-755, Apr.
2017. 3-D additive manufacturing (AM) offers unprecedented flexibility in
the realization of complicated 3-D structures. Polymer jetting is one of the
promising 3-D AM techniques that utilizes photosensitive polymers as the
build material and is capable of precisely printing electromagnetic (EM)
components up into the THz range. In this paper, important design and
implementation aspects of polymer-jetting-based 3-D-printed EM
components are discussed. A number of 3-D-printable polymer materials and
their broadband EM characterization from GHz to THz are introduced. Design
methodologies specific for 3-D-printed antennas and other EM components
are presented. As examples, various 3-D-printed devices operating from GHz
to THz frequency, including electromagnetic crystals (EMXT), waveguide,
horn antenna, gradient index (GRIN) lenses, as well as 3-D AM-enabled new
designs, such as millimeter wave (mmW)/THz, reflect array antennas,
computer-generated THz holograms, and so on are reviewed. Moreover,
current limitations and possible future improvements of the polymer jetting
technique for EM applications are discussed. This type of 3-D AM technique
is likely to enable many novel antenna and circuit architectures as well as
various interesting 3-D metamaterial structures.

• S. Koziel and S. Ogurtsov, “Multi-objective design of antennas using

19
 variable-fidelity simulations and surrogate models”, IEEE Trans.
Antennas Propag., vol. 61, no. 12, pp. 5931-5939, Dec. 2013. A
computationally-efficient procedure for multi-objective design of antenna
structures is presented. Our approach exploits the multi-objective
evolutionary algorithm (MOEA) working with a fast antenna surrogate model
obtained with kriging interpolation of coarse-discretization simulation data.
Response correction techniques are subsequently applied to refine the designs
obtained by MOEA. Our methodology allows us to obtain-at a low
computational cost-a set of designs corresponding to various trade-offs
between the antenna size and the refection coefficient. Two illustration
examples are considered: (i) an UWB monocone with two objectives being
reduction of the antenna size and minimization of the antenna reflection
coefficient in the bandwidth of interest, and (ii) a planar Yagi antenna with
the objectives being an increase of the end-fire gain and minimization of the
reflection coefficient, both in the bandwidth of interest.

• Y. Rahmat-Samii, J. M. Kovitz, and H. Rajagopalan, “Nature-inspired

optimization techniques in communication antenna design”, Proc. IEEE,
vol. 100, pp. 2132-2144, July 2012. This paper summarizes the primary
features inherent in current optimization methods typically applied to antenna
designs and demonstrates their effectiveness by applying particle swarm
optimization (PSO), a nature-inspired global optimization technique, to novel
antenna design solutions in wireless communications. The concept of the PSO
technique is briefly introduced and an outline of the important parameters that
are utilized is summarized. Next, an implementation strategy combining PSO
with numerical algorithms for electromagnetic solutions, namely the finite
element method (FEM) and the method of moments (MoM), is discussed. In
20
 both realizations (PSO–FEM and PSO–MoM), the PSO technique drives the
design variables, such as the antenna dimensions, geometrical features, etc.,
and the full-wave electromagnetic analysis engines evaluate the fitness
function for the optimizer. Optimized antenna designs including a multiband
handset antenna and an E-shaped patch antenna for circularly polarized (CP)
applications are presented. Measurement results of prototype optimized
designs are shown to demonstrate the functionality and effectiveness of the
methodologies presented in this paper.

• M. T. Hagan and M. B. Menhaj, “Training feedforward networks with the

Marquardt algorithm,” IEEE Trans. Neural Networks, vol. 5, no. 6, pp. 989-
993, 1994.The Marquardt algorithm for nonlinear least squares is presented
and is incorporated into the backpropagation algorithm for training
feedforward neural networks. The algorithm is tested on several function
approximation problems, and is compared with a conjugate gradient algorithm
and a variable learning rate algorithm. It is found that the Marquardt algorithm
is much more efficient than either of the other techniques when the network
contains no more than a few hundred weights

• T. Sallam, A. B. Abdel-Rahman, M. Alghoniemy, Z. Kawasaki, and T. Ushio,

“A neural-network-based beamformer for phased array weather radar,”
IEEE Trans. Geosci. Remote Sens., vol. 54, no. 9, pp. 5095-5104, Sept. 2016.
Phased array weather radars, particularly with high temporal resolution,
essentially need a robust and fast beam- former to accurately estimate
precipitation profiles such as re- flectivity and Doppler velocity. In this paper,
we introduce a neural-network-based beamformer to address this problem. In
particular, the optimum weight vector is computed by modeling the problem
21
 as a three-layer radial basis function neural network(RBFNN), which is
trained with I/O pairs obtained from the optimum Wiener solution. The
RBFNN was chosen because of its characteristic of accurate approximation
and good generalization, and its robustness against interference and noise. The
proposed RBFNN beamforming method is compared with traditional beam-
forming methods, namely, Fourier beamforming (FR), Capon beamforming,
and the flower pollination algorithm (FPA), which is a recently proposed
nature-inspired optimization algorithm. It is shown that the RBFNN approach
has nearly optimal performance in various precipitation radar signal
simulations relative to the rival methods. The validity of the RBFNN
beamformer is demonstrated by using real weather data collected by the
phased array radar (PAR) at Osaka University, and compared with, in addition
to the FR and FPA methods, the minimum mean square error beamforming
method. It is shown that the RBFNN method estimates the reflectivity of the
PAR at Osaka University with less clutter level than those of the other three
methods.

• S. B. Imandoust and M. Bolandraftar, “Application of K-Nearest Neighbor

(kNN) Approach for Predicting Economic Events: Theoretical
Background,” Int. J. of Eng. Research and Appl., vol. 3, no. 5, pp. 605- 610,
Sep-Oct 2013. In the present study k-Nearest Neighbor classification method,
have been studied for economic forecasting. Due to the effects of companies’
financial distress on stakeholders, financial distress prediction models have
been one of the most attractive areas in financial research. In recent years,
after the global financial crisis, the number of bankrupt companies has risen.
Since companies' financial distress is the first stage of bankruptcy, using
financial ratios for predicting financial distress have attracted too much
22
 attention of the academics as well as economic and financial institutions.
Although in recent years studies on predicting companies’ financial distress
in Iran have been increased, most efforts have exploited traditional statistical
methods; and just a few studies have used nonparametric methods. Recent
studies demonstrate this method is more capable than other methods.
Keywords: Predicting financial distress

ANSYS HFSS

3D Electromagnetic Field Simulator for RF and Wireless Design Ansys HFSS is

3D electromagnetic (EM) simulation software for designing and simulating high-
frequency electronic products such as antennas, antenna arrays, and RF or
microwave components, high speed interconnects filters, connectors, IC packages
and printed circuit boards. Engineers worldwide use Ansys HFSS to design high-
frequency, high-speed electronics found in communications systems, radar systems,
advanced driver assistance systems (ADAS), satellites, internet-of-things (IoT)
products and other high-speed RF and digital devices.HFSS (High Frequency
Structure Simulator) employs versatile solvers and an intuitive GUI to give you
unparalleled performance plus deep insight into all your 3D EM problems. Through
integration with Ansys thermal, structural and fluid dynamics tools, HFSS provides
a powerful and complete multiphysics analysis of electronic products, ensuring their
thermal and structural reliability. HFSS is synonymous with gold standard accuracy
and reliability for tackling 3D EM challenges by virtue of its automatic adaptive
meshing technique and sophisticated solvers, which can be accelerated through high
performance computing (HPC) technology.The Ansys HFSS simulation suite
consists of a comprehensive set of solvers to address diverse electromagnetic

23
problems ranging in detail and scale from passive IC components to extremely large-
scale EM analyses such as automotive radar scenes for ADAS systems. Its reliable
automatic adaptive mesh refinement lets you focus on the design instead of spending
time determining and creating the best mesh. This automation and guaranteed
accuracy differentiates HFSS from all other EM simulators, which require manual
user control and multiple solutions to ensure that the generated mesh is suitable and
accurate. With Ansys HFSS, the physics defines the mesh rather than the mesh
defining the physics.Ansys HFSS is the premier EM tool for R&D and virtual design
prototyping. It reduces design cycle time and boosts your product’s reliability and
performance. Beat the competition and capture your market with Ansys HFSS.
ANSYS simulation technology enables you to predict with confidence that your
products will thrive in the real world. Customers trust our software to help ensure
the integrity of their products and drive business success through innovation.
Industry Standard Full Wave, Electromagnetic Field Simulation HFSS sets the gold-
standard for accuracy, advanced solver and high-performance computing
technology, making it the ‘go to’ tool for engineers designing high-frequency and
high-speed electronics found in communication systems, radar systems, satellites,
smart phones and tablet devices. Rigorous Validation Sign-off quality high-
frequency EM results that allow customers to simulate and go straight to
manufacturing. With HFSS, engineers can extract parasitic parameters (S, Y, Z),
visualize 3D electromagnetic fields (near- and far-field) and generate Full-Wave
SPICE™ models that link to circuit simulations. Easy to Use, Versatile and Fast
Features such as automatic adaptive meshing, versatile design entry and advanced
high-performance computing technology put analyst-quality solvers in the hands of
the designer.

24
Ansys HFSS for Antenna Design

This application brief describes antenna design using Ansys HFSS, the industry
leading 3D Electromagnetic (EM) simulation tool for high frequency and high speed
electronic components. It highlights several antenna-related applications with
emphasis on antennas on or around other Structures. With multiple simulation
technologies and powerful automated adaptive mesh refinement Providing gold
standard accuracy, HFSS can help antenna designers who are constantly Challenged
with implementing designs across more and more frequency bands inside a smaller
And smaller footprint. With these additional technical challenges along with the ever
shrinking time to market, Simulation with HFSS is a must-have in the antenna design
and integration process.

Applications

The original ISM specifications envisioned that the bands would be used primarily
for non communication purposes, such as heating. The bands are still widely used
for these purposes. For many people, the most commonly encountered ISM device
is the home microwave oven operating at 2.45 GHz which uses microwaves to cook
food. Industrial heating is another big application area; such as induction heating,
microwave heat treating, plastic softening, and plastic welding processes. In medical
settings, shortwave and microwave diathermy machines use radio waves in the ISM
bands to apply deep heating to the body for relaxation and healing. More recently
hyperthermia uses microwaves to heat tissue to kill cancer cells [1] and [8].
However, as detailed below, the increasing congestion of the radio spectrum, the
increasing sophistication of microelectronics, and the attraction of unlicensed use,
has in recent decades led to an explosion of uses of these bands for short range
25
communication systems for wireless devices, which are now by far the largest uses
of these bands. These are sometimes called "non ISM" uses since they do not fall
under the originally envisioned "industrial", "scientific", and "medical" application
areas. One of the largest applications has been wireless networking(WiFi). The
bIEEE 802.11wireless networking protocols, the standards on which almost all
wireless systems are based, use the ISM bands. Virtually all laptops ,tablet
computers, computer printers and cellphones now have 802.11wireless modems
using the 2.4 and 5.7 GHz ISM bands. Bluetooth is another networking technology
using the 2.4 GHz band, which can be problematic given the probability of
interference. [9]Near field communication devices such as proximity cards and
contactless smart cards use the lower frequency 13 and 27 MHz ISM bands. Other
short range devices using the ISM bands are:wireless microphones ,monitors, garage
openers, wireless doorbells,keyless entry systems for vehicles,radio control
channels, RFID systems for merchandise, and tracking systems.

Some electrodeless designs are ISM devices, which use RF emissions to

excite fluorescent tubes. Sulfur lamp sare commercially available plasma lamps,
which use a 2.45 GHzmagnetronto heat sulfur into a brightly glowing plasma.

Long-distance wireless systems have been proposed and experimented with

which would use high-power transmitters and rectennas, in lieu of to send power to
remote locations.NASA has studied using microwave power on 2.45 GHz to send
energy collected by solar power satellites back to the ground.

Also in space applications, a Helicon Double Layerion thrusteris a prototype

spacecraft propulsion engine which uses a 13.56 MHz transmission to break down
and heat gas into plasma.
26
Applications in the ISM band include, wireless LANs, short range links for
advanced traveler systems (electronic toll collection), garage door openers, home
audio distribution, cordless phones, private point to point links, remote control,
wireless telemetric systems (e.g. electrical power consumption monitoring) etc.
Applications seem to be limited by the imagination rather than technology.

27
 CHAPTER 3

METHODOLOGY

➢ 3.1.1 Monopole Antenna

A monopole antenna is a class of radio antenna consisting of a straight rod-shaped

conductor, often mounted perpendicularly over some type of conductive surface,
called a ground plane.[1][2][3] The driving signal from the transmitter is applied, or
for receiving antennas the output signal to the receiver is taken, between the lower
end of the monopole and the ground plane. One side of the antenna feed line is
attached to the lower end of the monopole, and the other side is attached to the
ground plane, which is often the Earth. This contrasts with a dipole antenna which
consists of two identical rod conductors, with the signal from the transmitter applied
between the two halves of the antenna.

A typical mast radiator monopole antenna of an AM radio station in Chapel Hill,

North Carolina. The mast itself is connected to the transmitter and radiates the radio
waves. It is mounted on a ceramic insulator to isolate it from the ground. The other
terminal of the transmitter is connected to a ground system consisting of cables
buried under the field. The monopole is often used as a resonant antenna; the rod
functions as an open resonator for radio waves, oscillating with standing waves of
voltage and current along its length. Therefore the length of the antenna is
determined by the wavelength of the radio waves it is used with. The most common
form is the quarter-wave monopole, in which the antenna is approximately one
quarter of the wavelength of the radio waves. However in broadcasting monopole

28
antennas 5/8 = 0.625 wavelength long are also popular, because at this length a
monopole radiates a maximum amount of its power in horizontal directions. The
monopole antenna was invented in 1895 by radio pioneer Guglielmo Marconi; for
this reason it is sometimes called the Marconi antenna.[4][5][6] Common types of
monopole antenna are the whip, rubber ducky, helical, random wire, umbrella,
inverted-L and T-antenna, inverted-F, folded unipole antenna, mast radiator, and
ground plane antennas. The load impedance of the quarter-wave monopole is half
that of the dipole antenna or 37.5+j21.25 ohms.

Fig 3.1.1 Basic monopole antenna

Common types of monopole antenna are the whip, rubber ducky, helical, random
wire, umbrella, inverted-L and T-antenna, inverted-F, folded unipole antenna, mast
radiator, and ground plane antennas.

29
 ➢ 3.1.2 The Advantages of Monopole Antennas
1. As the monopole antenna gets longer and the ground losses are reduced, the
efficiency of the antenna gets better. Vertical monopole antennas can achieve
efficiencies of up to 80%.

2. A vertical monopole antenna can be used for any frequency shorter than two
thirds of the wavelength.

3. Monopole antennas are a simple omnidirectional antenna that takes up far less
space than an array of wheel antennas stacked on top of each other. The monopole
antenna can handle communications in any direction except straight up above the
antenna.

4. Monopole antennas are easy to build and install. Passive monopoles are cheap to
make and rugged.

5. If working with frequencies over 800 MHz, the monopole antenna can be made
out of the trace on a printed circuit board; this is standard for cell phones.

6. A monopole antenna has relatively high reactive impedance over most of its
frequency range. Put an active amplifier with ah high input impedance, and you
can transform the impedance without losing any sensitivity.

7. In the field of communication systems, whenever the need for wireless

communication arises, there occurs the necessity of an antenna. Antenna has the
30
capability of sending or receiving the electromagnetic waves for the sake of
communication, where you cannot expect to lay down a wiring system.

➢ 3.1.3 The Disadvantages of Monopole Antennas

Because you are radiating equally in all directions, you have equally poor radiation
in all directions.

The “torus” shape doesn’t extend to the top of the antenna because the voltage
increases as it travels up the antenna. In reality, the signal is sent from the bottom
two thirds of the antenna.

Metal objects and the ground itself can cause signal reflections, so you may get a
signal that is both horizontally and vertically polarized.

Inside a shielded chamber, monopole antennas can have impedances vary by

orders of magnitude, making it difficult to calibrate. Take measurements too close
to the tip of the antenna or ground plane, and they will likely be wrong.

Because you are radiating equally in all directions, you have equally poor radiation
in all directions. The “torus” shape doesn't extend to the top of the antenna because
the voltage increases as it travels up the antenna.

As the monopole antenna gets longer and the ground losses are reduced, the
efficiency of the antenna gets better. Vertical monopole antennas can achieve
efficiencies of up to 80%. A vertical monopole antenna can be used for any
frequency shorter than two thirds of the wavelength
31
3.1.4 APPLICATIONS
➢ Military Applications:
In the military sector, high-velocity aircraft, missiles rockets and spacecrafts need
low-profile and light weight antennas for conformal mounting outside their surfaces.
The micro strip antenna is best suited for this requirement, As these antenna would
neither 13 disturb the aerodynamic flow nor protrude inward to disrupt other already
crowded space. The uses of micro strip antennas in the military sector have been
numerous. These include functions such as altimeters, beacons,guidance fusing,
telemetry command, communication radar, ECM, ECCM, GPS and so on. [53] and
[54]

➢ Space Applications:
In the space sector, numerous applications of microstrip antennas have been
implemented. To name a few, the following space programs have used microstrip
antennas as arrays. Earth Limb Measurement Satellite (ELMS), International Sun
Earth Explorer (ISEE), SEASET, Shuttle Imaging Radar (SIR), GEOSTAR, Solar
Mesospheric Explorer (SME) and Mars Pathfinder. [54]

➢ Mobile System and Base System Applications:

Microstrip antennas are finding use for land-mobile base stations, maritime and
aeronautical mobile systems in present scenario. In a cellular mobile phone, a
service region enclosed by a base station antenna is separated into mini sectors to
raise the utilization of RF channels. These antennas are fabricated to have a zone or
multi-beam model to form three to six zones in 360o coverage. These patterns can
be synthesized using an antenna array and microstrip antennas can be used as the
most appropriate antenna elements.

32
3.1.5 Proposed antenna design parameters

To demonstrate the proposed idea, we use a reference double T-shaped monopole

antenna [13] (as shown in Fig. 1) as the example. The performance of this antenna
depends mainly on five design parameters l21, l22, w1, w2, w. In the design process,
we allow these five parameters to vary while keeping the other three parameters L,
h1 and h2 at their fixed values, as mentioned in [13]. These five geometric
parameters act as the explanatory variables or input variables for ML models. The
data is collected in batches and the corresponding R-squared values of the fitted
model trained by the accumulated data is monitored. As the sample size gets bigger,
the R-squared value first improves substantially but then levels off gradually. In our
numerical experiment, 450 samples give reasonably high and stable R-squared
values (larger than 0.85 for each case). Hence, N is chosen as 450 sample points and
data is collected by varying these five geometric parameters at different values using
random sampling. Theoretically, increasing the training samples improves the
prediction accuracy, as a larger number of sample points contain more information
about the underlying data generation scheme and therefore lead to better estimation
results. The same pattern is observed here when data is accumulated in batches.
Therefore, we choose sufficient number of training samples in batches, till the R-
squared becomes stable. The sample points are simulated using ANSYS high
frequency structure simulator (HFSS) [15] and antenna reflection coefficients in the
format of .s1p files are extracted from the simulations. Using each .s1p file,
performance of this antenna for each sample point is assessed by extracting figure
of merit (F OM), which is defined to obtain the maximal bandwidth in the two
desired bands of the antenna. In this study, F OM is defined as the sum of absolute
values of reflection coefficient (S11) in dB for frequency points in the range of 2.4-
3.0 GHz and 5.15-5.6 GHz. Basically, we add the absolute value of S11 at each of
33
the frequency point

Fig 3.1.2 Layout and design parameters of the referenced dual-band double – T shaped
monopole antenna

falling in the band of interest in order to calculate the FOM for a given design and
the same is represented mathematically

as follows:

where, f represents the frequency and S11(f) is the reflection coefficient value at that
frequency. While collecting the sample points, these parameters take values within

34
the following range of sample space, defined as: l21 ∈ [6.3, 7.3] , l22 ∈ [6.3, 7.3],
w1 ∈ [1, 3.5], w2 ∈ [1, 3.5], w ∈ [1, 3.5], with each parameter
taking a step size of 0.5 (all units are mm). In the antenna design process, these five
design parameters are input variables and FOM is the output or response variable.
The training data are represented by {(Xi , Yi), i = 1, 2, ..N}, where the input is X =
(l21, l22, w1,w2 , w) T ∈ χ or its transformation. The output Y is the value of figure
of merit (FOM). The goal is to learn a behavioral model based on the training set to
best describe the relationship between the expected F OM (F OM \ ) and the design
parameters. We represent this behavioral model by:
F OM \ = f (l21, l22, w1, w2, w) + , (2) where, is the error term and the function f is
a flexible mapping, which can be linear or nonlinear, continuous or discontinuous
function that is based on the main effects or including two-way interaction effects of
the input parameters. Machine learning methods are then used to search for the best
f ∈ F, a class of candidate models, to describe the relationship between X and Y and
make future predictions. To search for optimal design parameters based on the
obtained machine learning model, a very fine grid over the entire space χ with a step
size 0.1 mm (instead of 0.5 mm in the training set) is generated, consisting of a total
of 2,126,696 design points. The F OM values at all the design points are then
computed and the design parameter values that give the maximal F OM value are
identified. The performance of this analysis is verified by comparing the predicted
F OM at a set of test points with its actual value obtained from the HFSS simulation.
In the following subsections, the three machine learning techniques (namely, Least
absolute shrinkage and selection operator, Artificial neural network and k-nearest
neighbor) that have been used in this work to obtain the behavioral model f, are
explained.

35
Least Absolute Shrinkage and Selection Operator (lasso)

The lasso technique is a sparse regression and predictor selection algorithm that
estimates a linear model subject to some conditions. Basically it searches the value
of regression coefficients that can minimize the residual sum of squares, subject to
the condition that the absolute values of the re- gression coefficients are less than a
constant [16]. Denote the training data by (mk,nk), k = 1, 2, .., N, where mk = {mk1,
..., mkp} T
are p–dimensional predictor variables and nk are associated responses. The linear
estimation model predicts the response for a given input mk, as

nck = α+E p βpmkp. In a simple linear regression problem, the values of ( α, β) are

obtained by applying the least squares method i.e., minimizing

the difference between the actual response value nk and the estimated value (nck)

as defined below:

The standard regression analysis includes all the predictor variables in the model
fitting, irrespective of the magnitudes of their effects on the output. If some predictor
variables are not informative for prediction, these variables are regarded

36
as unimportant and should be removed from the final model, in order to improve
both prediction accuracy and model interpretability. This can be done by lasso,
which calculates the estimate (α, β) by minimizing the sum of residual squares
subject to the constraint Pp |βp| ≤ t. Here, t ≥ 0 is a tuning parameter, which controls
the amount of shrinkage that is applied to the estimates. In our analysis, we fit the
linear model by using the function lm() in R [17], which is a public-domain statistical
computing and graphics software. To increase model flexibility and capture any
nonlinear relationship between the predictors and the response, we have also
considered and compared various forms of nonlinear transformation of the raw
design parameters, including log, exponential, quadratic, cubic, etc., and finally
decided that a quadratic transformation is the most appropriate. To implement the
lasso procedure, we used the function lasso() in R. For this lasso model, the five
design parameters and their quadratic transformations are used as the predictor
variables and the F OM is used as the response variable. The optimal value of the
tuning parameter t is determined by the 5-fold cross validation method [18]. The
final fitted model predicted by the lasso is as follows:

37
 3.1.3 ANN Architecture based on MLP

Artificial Neural Network (ANN) Analysis

ANN (Artificial Neural Network) is a computational model that is inspired by the

function of biological neural net- works. ANN consists of a group of artificial
neurons, which processes information over interconnection. There are many
different ANN structures used in the literature. Multilayer perceptron’s (MLPs)
[19]–[21] which are successfully and commonly employed in engineering problems
are preferred in this study because of their ability to learn and model complex
relationships. The MLP can be trained by many algorithms such as Levenberg-
Marquardt (LM), back-propagation, delta- bar-delta and so on. In this work, we train
the MLPs using the Levenberg-Marquardt algorithm [22], which has the abilities of
fast learning and good convergence. For our problem setup, which has only five
parameters in the input space and needs a regression model, the LM algorithm is one
of the most efficient training algorithms for small and medium sized patterns. The
MLP consists of three layers: input layer, output layer and a hidden layer, as shown
in Fig. 2. Neurons in the input layer distributes the input signal ui , to the neurons in
the hidden layer. Each neuron j in the hidden layer adds up its input signals ui after
multiplying weights to each term depending on the respective connections wij from
the input layer and computes its output vj as a function g of the sum and a bias value,
bj added to it, i.e., where, g (.) can be a simple threshold function, a sigmoid,
hyperbolic tangent, a radial basis function, a purelin function, etc. The output of
neurons in the output layer is computed similarly. The simplest algorithm that can
be used is a
38
first-order error back-propagation (EBP) algorithm [23] which is one of the primitive
training algorithms for neural networks. However, the EBP algorithm suffers
because of its low training efficiency. This can be improved by using dynamic
learning rates or by using the second order algorithms, such as Newton algorithm
and Levenberg–Marquardt (LM) algorithm [24] which will help in increasing the
training speed of the EBP algorithm. The LM algorithm is a blend of vanilla gradient
descent and Gauss-Newton iterations, which gives it better convergence speed over
vanilla gradient descent algorithms. LM is also able to provide a solution for
nonlinear least squares minimization. Since it combines the EBP algorithm and
Newton algorithm, LM algorithm is considered to be most efficient algorithm for
small and median sized patterns. LM algorithm is implemented for training the
neural net- work for this study, using the neural network toolbox of MATLAB [25].
In our analysis, the data collected from .s1p file is divided into three parts: 70 % of
data is used for training and 15 % each used for testing and validation respectively.
The input layer consists of 5 design parameters. hidden layer of 5 hidden nodes and
output layer of single node for
F OM. .

39
C. k–nearest neighbor (kNN)

The k–nearest neighbor (kNN) [26] is an instance-based algorithm for supervised

learning, which defines the similarity between sample points and makes predictions
for new data points based on their similarity to the data that is already present in the
training set. The similarity measure is typically expressed by a distance measure
such as the Euclidean distance, cosine similarity or the Manhattan distance. For any
given new data point, the kNN algorithm first calculates its distance to all stored data
points, which are used to determine its k-nearest neighbors. Next, the outputs of
these neighbors are gathered to produce a weighted average, which will then be
assigned to the new data point. The weight of each neighbor is inversely proportional
to its distance to the target data point. In simple words, the nearer neighbors
contribute more to the average than the more distant one. To implement kNN in this
analysis, we first use 10-fold cross validation to choose an appropriate value of k. It
is observed that k=5 gives the minimum cross validation error and therefore 5 nearest
neighbors are chosen for each data point for prediction. We use the function,
knn.reg() in the FNN package of R [27] in order to obtain the kNN model for the
reference antenna. Hence, by using above mentioned three ML techniques we obtain
three separate behavioral models to relate design parameters to F OM of the
reference antenna. Next, these models are used to estimate performance of the
antenna at various design points in terms F OM and also the accuracy of the
estimations is tested by comparing results with HFSS. More details regarding this
have been done

40
 CHAPTER 4

RESULTS AND DISCUSSION

Using the behavioral models obtained from the artificial neural network model,
lasso model, and kNN model, respectively, we identify the optimal design
parameters which produce the highest F OM value. For this purpose, we predict the
F OM values for all possible combinations of values of the five design parameters
and locate the maximum F OM point. In particular, the five antenna design
parameters are varied in the following range: l21 ∈ [6.3, 7.3] , l22 ∈ [6.3, 7.3], w1 ∈
[1, 3.5], w2 ∈ [1, 3.5], w ∈ [1, 3.5], with each parameter having a step size of 0.1
(all units are in mm). simulation tool and summarized in Table I. The first column
of the top three rows in this table shows the design parameters value that is obtained
by each ML technique to give maximum F OM. The next three columns list the
corresponding maxi- mum F OM value predicted by each ML technique and the last
column is the F OM obtained from HFSS if the design parameters mentioned in the
first column are substituted in HFSS design. For comparison the design parameters
value of the reference antenna as mentioned in [13] are also stated in the first column
of the last row of Table I and in next columns of the fourth row, the corresponding
F OM value predicted by each technique as well as HFSS is shown. The fifth row in
this table mentions the total computation time corresponding to each of the ML
technique i.e. the time it took to train the ML model and then search the
corresponding best set of design parameters from that model. In the following
subsections, the results of Table I for each ML techniques are discussed in more
detail.

41
Table 4.1 Predicted FOM, design parameter values and computation time comparison

42
lasso Results

The lasso model is trained with 20 parameters in the input space, consists of five
original design parameters and their square as well as their cross product terms. Five-
fold cross validation [16] is used to determine the optimal value of the tuning
parameter. Though lasso is known to be able to produce sparse solutions, for this
case none of the coefficients becomes zero (as observed from Equation (4)), which
implies that all the predictors have non-trivial effects on the output performance. For
the model predicted by lasso as represented in Equation (4), the sensitivity analysis
is performed by perturbing few coefficients and the results are mentioned in Table
II. It can be observed from Table II that maximum value of F OM and the design
parameter values for which this maximum occurs for the perturbed model, remains
almost the same to the values predicted from the actual model. This proves the
robustness of the lasso model against slight random perturbations

43
Table 4.2 Sensitivity Analysis of predicted lasso model: Predicted FOM and design parameter values
when coefficients of Equation (4) are perturbed

44
The maximum F OM predicted by the lasso model is mentioned in first row of
Table I, which can be achieved by choosing the design parameter values as: l21 =
7.3 mm, l22 = 6.3 mm, w1 = 1 mm, w2 = 3.5 mm and w = 3.5mm. The trained
lasso model also predicts the F OM value at this design location equal to 298.56.
To verify the results, a HFSS simulation for these values of design parameters is
done. The corresponding value of F OM obtained from HFSS is 298.06

B. Artificial Neural Network Results

The results obtained from neural networks are more precise than lasso. For neural
network analysis, 70% of data was used for training and 15% each is used for testing
and validation, respectively. The input layer consists of five nodes, each representing
one design parameter, a hidden layer with 5 hidden nodes, and the output layer of
one single node, representing the F OM. As observed from the second row of Table
I, the maximum F OM is expected to occur at l21 = 7.3 mm, l22 = 6.3 mm, w1 = 1
mm, w2 = 3.5 mm and w = 3.5mm by the ANN model and the corresponding F OM
value is predicted to be equal to 297.68. The F OM value for these design parameters
is also checked through the HFSS simulation; the resulting value is 298.06. It can be
observed from here that the results obtained from lasso and neural network are very
close to each other.

C. k-Nearest Neighbor Results As mentioned earlier, for the k-Nearest neighbor

(kNN) analysis here, we use k = 5 and Euclidean distance based weighted average,
to estimate F OM for any design parameter set using its training data. As shown in
third row of Table I, the maximum F OM predicted by kNN is equal to 288.67, which
is given by the design parameters l21 = 7.3 mm, l22 = 6.3 mm, w1 = 1.2 mm, w2 =
45
3.3 mm and w = 3.1mm; for this particular set of values HFSS gives F OM = 280.53.
This implies a percentage error of 2.90%. Fig. 3 (a)-(e) shows the plots to compare
performance of the ML techniques. The values of F OM predicted by ML techniques
and obtained from HFSS, are plotted with respect to each of the five design
parameters. For each plot, only one design parameter is varied along the horizontal
axis, while the other four design parameters are kept as constant mentioned on the
top of each graph. For HFSS results, each design parameter is varied with a step size
of 0.5 mm because simulation for each design parameter set is quite time-consuming.
For the three ML techniques-predicted F OM, each design parameter is varied with
a step size of 0.1 mm because once the ML models relating the input and output
parameters have been established, F OM can be predicted much faster compared to
doing HFSS simulations. The vertical axis depicts the F OM values. The red curve
represents F OM from lasso prediction, black curve from ANN prediction, pink
curve from kNN prediction and blue curve is for F OM values calculated from HFSS
simulations. From Fig. 3, it is observed that the values from lasso and neural
networks are quite close to the actual F OM value obtained from HFSS, since the
red and black curves closely trace the blue curve in all the plots. This proves the
accuracy of lasso prediction and ANN prediction. On the other hand, the kNN
prediction (pink curve in Fig. 3) is little deviated from the actual values obtained
from HFSS and the other two ML techniques. Still, the kNN model is able to predict
dependence of F OM on each design parameter like the other models. One possible
reason for less precision in the kNN results can be due to its simple model structure,
which is sensitive to irrelevant or redundant features because all features contribute
while making prediction [26]. Also, the distribution of data and type of distance used
also affects the prediction values in the case of kNN. Therefore, these are factors that
may be responsible for the deviation of kNN prediction compared to the other two
techniques. The maximum F OM predicted by both lasso and ANN occurs for the
46
same set of design parameter values (l21 = 7.3 mm, l22 = 6.3 mm, w1 = 1 mm, w2
= 3.5 mm and w = 3.5 mm). These values are substituted in HFSS model and the
simulated results are compared with the results for design parameters set mentioned
in [13] in order to compare the bandwidth performance. Fig. 4 shows the S11 plot
for these two cases. It can be observed that for design parameters predicted by lasso
and ANN (red curve), larger bandwidth in the two bands is achieved compared to
that for the design parameters mentioned in [13] (blue curve). Hence, for this work
by using lasso and ANN modeling, we not only save time during optimization but
are also able to achieve better performance for the reference antenna structure. As
compared to the optimization done using EM solvers, ML optimization is quite fast.
This is clearly evident from the fifth row of Table I, as the computation time of each
algorithm is just a few hundreds of seconds which is significantly less than the time
taken by EM solver to simulate even one design point. If the same optimization has
to be done for these five design parameters by using an EM solver alone, it would
take a few hours. In addition, we can change the optimization goal later after data
collection by using ML optimization, while the same is not true for a standard
optimization by an EM solver alone. Also, optimization using an EM solver requires
designer’s empirical knowledge to choose the best solution, while ML optimization
automatically chooses the optimal design. The main advantage of the ML techniques
is its superior ability to solve large-scaled optimization problems. The successful
results for this work prove ML techniques to be a new and powerful tool to tackle
difficult design problems such as the initial design. These techniques can further be
used to solve even more complex problems. For example, one can allow a flexible
choice of the antenna shape, structure, and material, by formulating a large
optimization problem, and then use the ML to search within a high-dimensional
input space for optimal design. To achieve this, we will need to collect more training
data.
47
4.1 RESULTS

Predicted as well as simulated F OM values with respect to change in (a) l21, (b)
l22, (c) w1, (d) w2 and (e) w respectively while keeping other four constants to the
values shown on top of each plot and legend shown on bottom-left.

(a) FOM values with respect to change in l21

48
(b) FOM value with respect to change in l

49
 (C) FOM values with respect to change in w1

(d) FOM values with respect to change in w2

50
 (e) FOM values with respect to change in w3

(f) Simulated S11 versus frequency response for the

design parameters values mentioned in [13] (blue curve) and
for design parameters values predicted by lasso and ANN (red
curve) in order to attain maximum F OM.

51
 CHAPTER 5

CONCLUSION

In this work three machine learning techniques namely least absolute shrinkage and
selection operator (lasso), artificial neural network (ANN) and k-nearest neighbors
(kNN) are used to automatically identify the optimal values of the design parameters
for a reference antenna where it can provide the best performance in terms of
bandwidth of two bands. The brief description about these techniques is first
presented in this paper and then how these techniques are applied to a reference
double T-shaped monopole antenna is explained. With the help of these ML
techniques, performance of the reference antenna is analyzed for 2126696 design
points within a few seconds by learning from the training dataset of 450 data points
only. Compared to kNN, both neural network and lasso give more accurate
predictions in our study. In summary, these new methods are much more efficient
than traditional method of EM simulation optimization for achieving optimal
antenna design. The results obtained from this research imply that ML techniques
have the power to revolutionize EM simulation technology. Due to computational
power limits of EM tools it is challenging and time consuming to optimize complex
antenna designs like 3-D antenna structures involving a large number of design
parameters. This problem can be addressed by incorporating ML techniques in
simulation tools. The ultimate goal of this work is to further generalize machine
learning methods for complex design structures such as structures manufactured by
3D printing technology. For example, 3D printing enables designs with large
numbers of degree of freedom and therefore optimizing all parameters through EM
simulations is both tedious and computationally intensive.

52
Our preliminary results have shown that machine learning techniques may be able
to enable versatile and potentially automated design of antennas which will be
beneficial for anumber of applications including internet of things

53
REFERENCES

[1] H. Xin and M. Liang, “3-D printed microwave and THz devices using polymer jetting
techniques,” Proc. IEEE, vol. 105, no. 4, pp. 737-755, Apr. 2017.

[2] T. Hastie and P. Simard, “Metrics and Models for Handwritten Character Recognition,”
Statistical Science, vol. 13, no. 1, pp. 54-65, 1998.

[3] T. S. Furey, N. Cristianini, N. Duffy, D. W. Bednarski, M. Schummer, and D. Haussler,

“Support vector machine classification and validation of cancer tissue samples using microarray
expression data,” Bioinformatics, vol. 16, no. 10, pp. 906-914, Oct. 2000.

[4] S. Koziel and S. Ogurtsov, “Multi-objective design of antennas using variable-fidelity

simulations and surrogate models, IEEE Trans. Antennas Propag., vol. 61, no. 12, pp. 5931-5939,
Dec. 2013.

[5] D. Z. Zhu, P. L. Werner, and D. H. Werner, “Design and Optimization of 3D Frequency

Selective Surfaces Based on a Multi-Objective Lazy Ant Colony Optimization Algorithm,” IEEE
Trans. Antennas Propag., vol. 65, no. 12, pp. 7137-7149, Dec. 2017.

[6] Y. Rahmat-Samii, J. M. Kovitz, and H. Rajagopalan, “Nature-inspired optimization

techniques in communication antenna design, Proc. IEEE, vol. 100, pp. 2132-2144, July 2012.

[7] A. Massa, G. Oliveri, M. Salucci, N. Anselmi, and P. Rocca, “Learning- by-examples

techniques as applied to electromagnetics,” J. of Electro- magn. Waves and Applications, pp. 1-
16, 2017.

[8] Z. Zheng, X. Chen and K. Huang, “Application of support vector machines to the antenna
design,” Int. J. RF Microw. Computer-Aided Eng., vol. 21, no. 1, pp. 85-90, 2010

[9] M. M.-Ramon, and C. Christodoulou, “Support Vector Machines for Antenna Array
Processing and Electromagnetics,” Morgan Claypool Publishers, 2006.

[10] T. Sallam, A. B. Abdel-Rahman, M. Alghoniemy, Z. Kawasaki, and T. Ushio, “A neural-

network-based beamformer for phased array weather radar,” IEEE Trans. Geosci. Remote Sens.,
vol. 54, no. 9, pp. 5095-5104, Sept. 2016.

[11] K. C. Lee, and T. N. Lin, “Application of neural networks to analyses of nonlinearly loaded
54
antenna arrays including mutual coupling effects,” IEEE Trans Antennas Propag., vol. 53, no. 3,
pp. 11261132, 2005.

[12] G. Dadashzadeh, M. Kargar, Y. Torabi, and B. Rahmati, “Broad-Band and Wide Scan Phased
Array Element Design Using Data Mining,” Appl. Comput. Electromagn. Soc. J., vol. 31, no. 3,
pp. 244-251, 2016.

[13] Y. L. Kuo and K. L. Wong, “Printed double-T monopole antenna for 2.4/5.2 GHz dual-band
WLAN operations,” IEEE Trans. Antennas Propag., vol. 51, no. 9, pp. 2187-2192, Sept. 2003.

[14] Y. Sharma, J. Wu, H. Xin, and H. H. Zhang , “Sparse linear regression for optimizing design
parameters of double T-shaped monopole antennas,” 2017 IEEE International Symposium on
Antennas and Propagation & USNC/URSI National Radio Science Meeting, San Diego, CA,
2017, pp. 347-348.

[15] ANSYS Electromagnetics Suite 15.0, ANSYS, Inc

.
[16] R. Tibshirani, “Regression shrinkage and selection via the lasso: a retrospective,” J. R. Stat.
Soc. Ser. B Stat. Methodol., vol. 73, no. 3, pp. 273-282, Jun. 2011.

[17] R: A Language and Environment for Statistical Computing, R Foundation for Statistical
Computing, Vienna, Austria, 2013.

[18] T. Hastie, R. Tibshirani, and J. Friedman, “The elements of statistical learning: data mining,
inference and prediction,” Springer, 2nd ed., 2016.

[19] I. Develi, “Application of multilayer perceptron networks to laser diode nonlinearity

determination for radio-over-fibre mobile communications,” Microw. Opt. Technol. Lett., vol.
42, no. 5, pp. 425-427, Sept. 2004.

[20] S. S. Haykin, “Neural networks: a comprehensive foundation,” Macmil-lan, 1994.

[21] J. -S. R. Jang, “Self-learning fuzzy controllers based on temporal back- propagation,” IEEE
Trans. Neural Networks, vol. 3, no. 5, pp. 714-723,1992.

[22] M. T. Hagan and M. B. Menhaj, “Training feedforward networks with the Marquardt
algorithm,” IEEE Trans. Neural Networks, vol. 5, no. 6, pp. 989-993, 1994.

[23] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-

55
propagating errors,” Nature, vol. 323, no. 6088, pp. 533- 536, Oct. 1986.

[24] J. -M. Wu, “Multilayer Potts Perceptrons with Levenberg-Marquardt Learning,” IEEE Trans.
Neural Networks, vol. 19, no. 12, pp. 2032-2043, Dec. 2008.

[25] H. Demuth and H.M. Beale, “Neural Network Toolbox for Use with Matlab– User’s Guide
Verion 3.0,” 1998.

[26] S. B. Imandoust and M. Bolandraftar, “Application of K-Nearest Neighbor (kNN) Approach

for Predicting Economic Events: Theoretical Background,” Int. J. of Eng. Research and Appl.,
vol. 3, no. 5, pp. 605- 610, Sep-Oct 2013.

[27] A. Beygelzimer, S. Kakadet, J.Langford, S. Arya, D. Mount and S. Li, “Fast Nearest
Neighbor Search Algorithms and Applications,” R package, version1.1.2.1,20

56