You are on page 1of 12

Received: 26 April 2021 Revised: 5 July 2021 Accepted: 20 August 2021

DOI: 10.1002/jnm.2951

RESEARCH ARTICLE

An optimized surrogate model using differential evolution


algorithm for computing parameters of antennas

Deniz Ustun1 | Feyza Toktas2 | Abdurrahim Toktas3

1
Department of Computer Engineering,
Engineering Faculty, Tarsus University,
Abstract
Tarsus, Turkey In this study, a method based on surrogate model (SM) for computational anal-
2
Department of Computer Engineering, ysis of antenna parameters such as the resonant frequency (RF) and band-
Engineering Faculty, Mersin University,
width (BW) is presented. Moreover, it is attempted to optimize the SM using
Mersin, Turkey
3 evolutionary optimization algorithms in order to further improve the accuracy
Department of Electrical and Electronics
Engineering, Karamanoglu Mehmetbey of the SM. In the conventional computational approaches, the weighting vec-
University, Karaman, Turkey tors of the SM have been analytically determined. We have optimally achieved
Correspondence
the weighting vectors of the SM through differential evolution (DE) and parti-
Abdurrahim Toktas, Department of cle swarm optimization (PSO) algorithms. The capabilities of the algorithms
Electrical and Electronics Engineering, are hereby compared with each other. The methodology is applied to the anal-
Karamanoglu Mehmetbey University,
70200 Karaman, Turkey. ysis of rectangular microstrip antenna (RMA), including a number of 33 mea-
Email: atoktas@kmu.edu.tr sured RMAs with different geometrical and electrical parameters. From the
total number of RMAs, 27 and 6 RMAs are, respectively, used in the construc-
tion and the test of the SM. Furthermore, the SM is verified through a compar-
ison with the literature in terms of total absolute errors (TAEs). The results
show that the SM with DE computes the most accurate RF and BW with the
TAEs of 0.0099 GHz and 0.131%, respectively. The accuracy of the SM is fur-
ther raised by 78%, thanks to the optimization of SM with DE. Therefore, a
novel computational analysis method based on SM is implemented to compu-
tation of an antenna parameter with higher accuracy, and SM is successfully
optimized by DE. The proposed method is able to easily implement to the
stringent engineering problems based on simulated or measured data for
computer-aided design (CAD).

KEYWORDS
bandwidth, differential evolution, microstrip antennas, particle swarm optimization,
resonant frequency, surrogate model

1 | INTRODUCTION

The antennas are the most essential air-interface elements of the wireless communication network. Adaptation of the
microstrip technology to the antenna field paves the way in modern communication systems, especially for mobile ter-
minals. Therefore, the native benefits of the microstrip like being light-weight, low profile, inexpensive, and easy
manufacturing are inhered to the microstrip antennas. They have been widely used in a broad area of applications from
wireless communications to biomedical technologies.1–3 Therefore, the analysis and design of the microstrip antennas

Int J Numer Model. 2021;e2951. wileyonlinelibrary.com/journal/jnm © 2021 John Wiley & Sons Ltd. 1 of 12
https://doi.org/10.1002/jnm.2951
2 of 12 USTUN ET AL.

have become crucial research topics. Since the antennas can merely exhibit high performance around the resonant fre-
quency (RF) as allowed by the range of bandwidth (BW), most of the studied characteristics in the analysis of an
antenna are the RF and BW. These characteristics highly associate with the geometrical and electrical parameters of a
microstrip antenna. Due to having a simple shape and low parameters, rectangular microstrip antennas (RMAs) have
been the most interested and investigated antenna geometry. Various methods for determining the characteristics of
RMAs such as the RF and the BW have been reported in the literature.4–20 The suggested methods mainly depend on
formulaic or neurocomputational models. Note that those approaches were commonly conducted on a dataset of mea-
sured RMAs formerly reported in references 4–6. For determination of the RF, several formulaic models based on edge
extension length of the transmission line model (TLM) have been fitted with the measured data,6–9 and many neuro-
models depending on artificial neural network (ANN) and adaptive network fuzzy inference system (ANFIS) have been
reported in references 10–17. For finding the BW, similar expressions built on the quality factor of the RMA have
been proposed in references 4,5,18 and ANN and ANFIS-based neuro-models have been developed in references
11,19,20. Concisely, the formulaic methods are more useful and operable, since they are simple and easy to use. How-
ever, their accuracy is comparatively low. That is why they were derived by using a limited number of fitting parame-
ters. On the contrary, the neurocomputational methods give more accurate and close results, however, they are difficult
to implement and test as compared with formulaic models. It is evident that there has been still a need for not only a
simpler but also more accurate model.
Surrogate model (SM), that is, an efficient and versatile computational technique to model complex systems of simu-
lation or experiment, has been developed in recent years.21–25 The SM utilizes a compact analytic framework, which
simulates the input and output patterns of a system. It is constructed on an interpolating mathematical function by
fitting the analytic model using weighting vectors in accordance with the pattern among the inputs and outputs. Hence,
fitting appropriately the analytic model is a key point for a tight and accurate SM. On the other hand, evolutionary opti-
mization algorithms have been successfully applied to a variety of engineering problems.26–28 There have been plenty of
varieties of nature-inspired evolutionary optimization algorithms. Among them, the well-known algorithms are ant col-
ony algorithm (ACO) mimicking the communication and direction-finding behavior of the ants,29 particle swarm opti-
mization (PSO) simulating the flock action of moving organisms, birds, and fishes,30 genetic algorithm (GA)31 and
differential evolution (DE)32 both modeling the mutation phenomenon of living organism, whale optimization algo-
rithm (WOA) mimicking humpback whale hunting action,33 artificial bee colony (ABC) imitating the collective forag-
ing of the honey bees,34 and butterfly optimization algorithm (BOA) emulating foraging phenomenon of the real
butterflies, which was developed inspiring by the natural evolution of the living organism species.26 DE, which is a sim-
ple and powerful algorithm, has been successfully applied to many expensive optimization problems in various fields,
and thus comes into prominence, thanks to its robustness and fast convergence. It is a promising idea that SM can be
optimized through an evolutionary optimization algorithm for providing a better-fitted model.
In this study, a novel radial basis function (RBF)-based SM for computing the performance parameters of the anten-
nas is presented through a measured RMAs dataset in terms of the RF and BW. The SM is optimized with the help of
the DE algorithm for determining the weighting vectors.35 Moreover, the performance of DE is corroborated by compar-
ing the results with PSO. The SM is well built and evaluated on the measured dataset of 33 RMAs.4–6 The RF and BW
of the SM are compared with the SM without DE, PSO as well as the best results of the neuro-models such as ANN16,19
and ANFIS,15,20 and formulaic models4–9 with regard to total absolute error (TAE). The SM with DE estimates the RF
with the most accurate TAE of 0.0099 GHz as compared with the TAE of 0.0742, 0.0245, 0.0113, 0.031, and 1.001 GHz
computed by the SM without DE, PSO, ANN, ANFIS, and the formulaic model, respectively. While the SM without DE,
PSO ANN, ANFIS, and the formulaic model, respectively, compute the BW with TAE of 0.589%, 0.328%, 0.969%,
0.229%, and 23.920%; the SM with DE achieves it with the most correct TAE of 0.131%. Consequently, the contribution
can be briefly manifested as follows:

• A computational method based on SM is introduced for computing the parameters of microstrip antennas.
• The proposed method is successfully implemented to the measured data for analysis of RMA regarding RF and BW.
• A novel method based on evolutionary optimization is proposed for achieving a more accurate SM rather than the
conventional approaches in which the weighting vectors of SM were derived analytically.
• The SM is elaborately compared with the-state-of-the-art formulaic model and neuro-models such as ANN and
ANFIS. The SM is the most accurate model in the comparison, as well as the accuracy of the SM, is further raised by
78%, thanks to the optimization of SM with DE.
• The performance of DE is also verified through a comparison with PSO.
USTUN ET AL. 3 of 12

It is worth noting that the proposed conceptual scheme can be also implemented to the tough engineering problem
with simulated parameters for computer-aided design (CAD). Note that the dataset in the study has two outputs RF
and BW. If the data had included other parameters such as gain and efficiency, the SM would also have computed these
parameters.

2 | THE GEOMETRY OF THE RMA A ND THE DATASET

A 3D geometry of RMA is depicted in Figure 1. The conventional RMA consists of a rectangular radiating patch with
W  L and a full ground plane, wherein between a substrate with a relative permittivity of εr and thickness of h is
embedded. The rectangular patch is energized by a coaxial feed, and the point of a feed is offset toward the x-axis from
the short edge of the RMA on the centerline for impedance matching.
Table 1 includes a dataset that comprises a number of 33 measured RMAs4–6 with various geometrical and electrical
parameters, which are used to build and evaluate the SM. These antenna parameters were also utilized in the previous
studies.4–20 Therefore, the results of the proposed SM can be fairly compared with those studies. As the same with the
previous study, the number of 27 RMA are used in the construction, the remainders (with bold type in Table 1) 6 are
utilized in the test of SM.

3 | THE EXISTING COMPUTATIONAL APPROACHES

Mainly two types of approaches that are neuro-models based on ANN or ANFIS and formulaic models have been
suggested for the computation of RF and BW in the literature.4–20 The schematic representations of those models are
depicted in Figure 2. The simplest one is the formulaic model given in Figure 2A derived empirically from the RMA
using the TLM. According to the TLM, the long edge of the antenna is supposed to extend due to the fringing effect at
the edges. Similarly, the formulaic models for the computation of BW have been derived upon the quality factor calcu-
lation.4,5,18 However, since they are in the form of a mathematical function, including input parameters and limited
coefficients, their nonlinear fitting ability becomes restricted. In the derivation of some models,8,9 evolutionary optimi-
zation algorithms were also employed to obtain more fitted models. Hence, although the formulaic models have easy
usage, their accuracy is relatively poor.
Neuro-models11,19,20 given in Figure 2B,C have been used for the computation of RF and BW, which consist of
many neuron layers with activation functions and weighting connections between layers. To construct the framework
of neuro-model, hyperparameters such as the number of layers, neurons, rules, and activation functions should be
properly determined. Moreover, the transferring and reusing of a trained model are relatively difficult. That is why
neuro-models are sometimes called closed box framework. Yet neuro-models' nonlinear fitting capability is higher than

FIGURE 1 The rectangular microstrip antenna (RMA) geometry


4 of 12 USTUN ET AL.

TABLE 1 Dataset of measured rectangular microstrip antennas (RMAs) for construction/training and test (dimension unit: mm)4–6

The parameters of the measured RMAs4–6

Antenna no. W L A h εr
1 8.50 12.90 4.15 0.17 2.22
2 20.00 25.0 6.83 0.79 2.22
3 10.63 11.83 3.90 0.79 2.22
a a a a a
4 20.74 26.2 10.2 0.79 2.55a
4b 11.85b 7.90b 4.10b 0.17b 2.22b
5 9.10 10.00 3.75 1.27 10.2
6 17.20 18.60 5.94 1.57 2.33
c
7 18.10 19.60 6.27 1.57 2.33
8 12.70 13.50 4.25 1.63 2.55
9 15.00 16.21 5.28 1.63 2.55
c
10 13.37 14.12 4.75 2.00 2.55
11 11.20 12.00 4.25 2.42 2.55
12 14.03 14.85 4.60 2.52 2.55
13 15.30 16.30 4.70 3.00 2.50
c
14 9.05 10.18 3.70 3.00 2.50
15 11.70 12.80 3.40 3.00 2.50
16 13.75 15.80 5.82 4.76 2.55
17 7.76 10.80 4.03 3.30 2.55
18c 7.90 12.55 3.00 4.00 2.55
19 9.87 14.50 3.75 4.50 2.55
20 10.00 15.20 3.45 4.76 2.55
21 8.14 14.40 3.10 4.76 2.55
22 7.90 16.20 3.50 5.50 2.55
23 12.00 19.70 2.55 6.26 2.55
24 7.83 23.00 4.25 8.45 2.55
25 12.56 27.56 3.20 9.52 2.55
26 9.74 26.20 3.10 9.52 2.55
c
27 10.20 26.40 3.55 9.52 2.55
28 8.83 26.76 3.06 10.00 2.55
29 7.77 28.35 3.20 11.00 2.55
30 9.20 31.30 3.00 12.00 2.55
31 10.30 33.80 3.60 12.81 2.55
32 12.65 35.00 3.70 12.81 2.55
c
33 10.80 34.00 3.70 12.81 2.55
a
For only computation of bandwidth.5
b
For only computation of resonant frequency.6
c
For the test.

the formulaic models due to multiple layers and numerous weighting connections. Their accuracy is, therefore, better
than the functional models. However, the construction task of models is complex as it needs a training phase, which
can be seen from the schematic views of neuro-models. Hence, while the formulaic models can be used with a simple
calculator, the neuro-models utilize computers as well as the training phase. On the other side, SM has an open box
framework raised on analytical expression. After the construction, it can be freely carried and utilized. In this perspec-
tive, it resembles the formulaic models. It is able to tightly fit any system, thanks to its high nonlinear behavior.
USTUN ET AL. 5 of 12

F I G U R E 2 Schematic representation of the models: (A) formulaic, (B) artificial neural network (ANN), and (C) adaptive network fuzzy
inference system (ANFIS)

4 | THE A NALYTICALLY CONSTRUCTION OF THE SM

The SM is constructed using an input–output interrelationship by fitting interpolating function, which is preferably fast
and reasonably accurate metamodel. Schematic block visual of SM is illustrated in Figure 3. An outcome of any input
can be approximately calculated via the SM for the test phase, which was not used in the raising of the SM.
Selections of the function and determining the weighting vectors are hence crucial in the construction of an
SM. Several interpolating functions, the well-knows are RBFs,36 response surface method (RSM),37 and Kriging,38 have
been proposed in the literature. RBF is widely used in the construction of SM due to its simplicity and being computa-
tionally cheap.39 Therefore, RBF is preferred in this study after trying the other function as well. The main structure of
the radial basis SM is given below:
XN
T
b
y ðb
xÞ ¼ λ ϕ ðr Þ þ β b
x þ α, ð1Þ
i¼1 i

where by ðbx Þ is the output vector of the SM for input vector xb. In our problem, x
b is the antenna parameter of the RMA
and b
y ðb
x Þ is the respective RF or BW. Each xbi refers to each antenna in our dataset of the RMA. λi ¼ ½λ1 , λ2 ,  , λN T is the
primary weighting vector with the number of input data N, that is, N ¼ 27, which is the number of RMAs in the
6 of 12 USTUN ET AL.

FIGURE 3 Schematic representation of the surrogate model (SM)

construction of the SM. βi ¼ ½β1 , β2 , , βd  is the secondary weighting vector and d is the dimension of the dataset, that
is, d ¼ 5, which is the number of antenna parameters. α is the weighting constant. r ¼ k b xl  b
x v k denotes the Euclidean
norm with l, v ¼ 1, 2,  , N. The following power-exponential RBF function is chosen in our implementation.
 p
r
ϕðr Þ ¼ exp : ð2Þ
σp

Here, σ and p are, respectively, the width and the power factors of the RBF, which allow tuning of the SM by setting
them manually. Hence, these factors should be determined as optimal as possible. After the input–output relationship
is constituted as above Equation (1), it is time to solve the following expression to find the weighting vectors and the
constant.36
    
ΦP λ b
y
¼ , ð3Þ
PT 0 c 0

h iT
T T
where P ¼ bx 1 ,  ,b
x N ; 1, …, 1 and C ¼ ½β1 , β2 , , βd , α . Therefore, the primary and secondary weighting vectors λ, β,
and constant α are analytically solved as follows:

   1  
λ ΦP b
y
¼ : ð4Þ
c PT 0 0

The final SM is thus constructed by instituting the determined weighting vectors into the main SM structure given
in Equation (1). Therefore, the RF or BW of any RMA can be computed through the constructed SM for the test phase.
In general, the weighting vectors are analytically determined as aforementioned. In this way, however, the model
might not be fitted tightly to the input–output interrelationship. In this study, for the first time, evolutionary optimiza-
tion algorithms such as DE and PSO are proposed to optimally find the weighting vectors together with the weighting
factors of the RBF.

5 | THE O PTIMIZATION O F SM WITH DE

A flowchart regarding the construction and optimization of SM with DE is illustrated in Figure 4. Accordingly, a
dataset, including both training and testing sets is first prepared to be input to the SM. The SM is then analytically con-
structed as described in the above section. Afterward, the DE is started for the optimization of SM so as to determine
the decision variables, which are the weighting vectors and factors λ, β, α, p, σ:
DE algorithm was developed inspiring by the natural evolution of the living organism species.35 DE mimics the evo-
lution phenomena mainly in three phases: mutation, crossover, and selection as seen from Figure 4. Assuming that the
evolution progresses from generation to generation. These three phases sequentially continue up to a predefined maxi-
mum number of generations Gmax. Given that chromosomes, which contain genes, are the structures within a cell of a
USTUN ET AL. 7 of 12

FIGURE 4 Flowchart of optimization of surrogate model (SM) with differential evolution (DE)

living organism. The population in DE accordingly comprises chromosomes each consisting of genes. The chromosome
stands for the decision variable referred to as the candidate solutions and the gene is its dimension.
In the initial phase, random values related to the candidate solutions for the optimization are generated (step i)
ðlÞ ðuÞ
between two specified lower bound (Lb) x j and upper bound (Ub) x j to constitute the initial generation (G) as
follows:
 
ðlÞ ðuÞ ðlÞ
8i ≤ NP ^ 8j ≤ D : x j,i,G¼0 ¼ x j þ randj ½0, 1: x j  x j , ð5Þ

where NP is the number of populations (chromosome number i ¼ 1, 2,  ,NP) and D is the dimension of decision vari-
ables (gene number j = 1, 2, …, D). The objective value of each generated decision variable is then calculated to assess
the quality of each chromosome. In the mutation phase, new variables are regenerated (step ii) by mutating the ran-
domly selected three individuals as allowed by mutation factor (F). The following operator is used to regenerate the
new variables.

8j ≤ D : nj,i,Gþ1 ¼ x j,r3 ,G þ F: x j,r 1 ,G  x j,r2 ,G : ð6Þ

Here, r1,2,3 is a randomly selected chromosome for mutated individuals nj,i,G + 1. All individuals are thus modified
by involving the two other variables into one, which belongs to randomly selected three individuals. In the crossover
phase, the mutated variables are randomly exchanged (step iii) with respective unmutated ones deciding by the follow-
ing operator, including a crossover rate (CR).
8 of 12 USTUN ET AL.

W
x j,n,Gþ1 if rand½0, 1 ≤ CR j ¼ jrand
8j ≤ D : x j,u,Gþ1 ¼ : ð7Þ
x j,i,G otherwise

In the selection phase committed by Equation (8), the variables, which make the objective function closer to the tar-
get are selected. In other words, the best variables are transferred to the subsequent generation, the others are deleted
(step iv).

x u,Gþ1 if f ðx u,Gþ1 Þ ≤ f ðx i,G Þ


8i ≤ NP : x i,Gþ1 ¼ : ð8Þ
x i,G otherwise

After the DE reaches the stopping criteria Gmax (step v), the final best decision variable λ, β, α,p, σ achieved so far are
substituted in the SM for accomplishing the optimized SM for computation of RF and BW of the RMAs. The DE
implemented to our problem is set by the control parameters given in Table 2.

6 | T H E CO M P A R A T I V E R E S U L T S

The RF and BW results of the analytically constructed SM and the SM optimized with DE are listed in Tables 3 and 4,
respectively. The RF results are even compared with those neuro-models ANN,16 ANFIS,15 and a formulaic expression9
in Table 3. Likewise, the BW results are compared with that of ANN,19 ANFIS,20 and a formulaic model.4 Moreover,
SM is optimized with PSO in order to corroborate the performance of DE through a comparison. Although more related
studies have been reported in the literature4–20 as mentioned above, those that have the best accuracy by ANN, ANFIS,
and a formulaic model are considered in the comparison for the sake of clarity. In order to concisely examine the
results, they are compiled in terms of the TAE regarding the construction (C)/training (T) and test at the bottom of the
tables.
From Table 3, the SM optimized with DE and PSO, respectively, computes the RF with TAEs of 0.0099 and
0.0245 GHz; on the other hand, ANN,16 ANFIS,15 and the formula9 estimate with the TAE of 0.0113, 0.031, and
1.001 GHz, respectively. Notice that the TAE of SM is improved from 0.0742 to 0.0099 GHz and 0.0245 GHz, thanks to
DE and PSO, respectively. The accuracy of the SM is hereby increased by approximately 87% and 67% by optimizing
with DE and PSO. From Table 4, the SM optimized with DE and PSO calculates the BW with TAE of 0.131% and
0.328% as compared with those TAEs are 0.969%, 0.229%, and 23.920% computed by ANN,19 ANFIS,20 and the formula.4
The TAE of SM is enhanced from 0.589% to 0.131% and 0.328%, thanks to DE and PSO, respectively. By optimizing SM
with DE and PSO, the accuracy of the SM is hence raised by 78% and 44%, respectively. Moreover, the SM optimized
with DE is the best among the others in the construction for computation of both the RF and BW.
It should be noticed that the proposed SM as similar to the existing computational approaches based on neuro and
formulaic models needs construction or training data to compute analysis parameters in the test process. Yet the results
hereby achieved show that the SM is more accurate than the neuro-models despite the fact that they have a closed box
framework having deeper complexity. Although the complexity of SM may be considered between the neuro and for-
mulaic models, the SM has both the precision of the neuro-models and the simplicity and handy usage of the formulaic
models. Since the SM is an open box analytic model, its reconstruction is simple and easy and hence it can be trans-
ferred and reused again. It is also demonstrated that evolutionary optimization algorithms such as DE and PSO can be
successfully applied to the optimization of SMs. It can be concluded that the SM is a useful, accurate, and fast technique
for the computational model of the antennas through CAD.

TABLE 2 Control parameters for setting the DE algorithm

Gmax NP Dimension F CR
250 30 70 0.6 0.8
Bound Lbλ,β Ubλ,β Lbα,p,σ Ubα,p,σ
Value 0 10 250 250
USTUN ET AL. 9 of 12

TABLE 3 Comparative results of the measured and computed RF with the literature

RF (GHz)

This study Literature

Optimized Analytic Neuro-model Formula


4,6 16 15
Antenna no. Measured SM with DE SM with PSO SM ANN ANFIS 9
1 7.740 7.740 7.740 7.740 7.740 7.740 7.737
2 3.970 3.970 3.970 3.970 3.970 3.970 3.925
3 7.730 7.730 7.730 7.730 7.730 7.730 7.730
4 8.450 8.450 8.450 8.450 8.451 8.450 8.415
5 4.600 4.600 4.600 4.600 4.600 4.600 4.600
6 5.060 5.060 5.060 5.060 5.060 5.056 4.948
7a 4.805 4.805 4.811 4.830 4.806 4.805 4.715
8 6.560 6.560 6.560 6.560 6.560 6.560 6.578
9 5.600 5.600 5.600 5.600 5.600 5.600 5.546
a
10 6.200 6.2037 6.1957 6.191 6.201 6.200 6.207
11 7.050 7.050 7.050 7.050 7.050 7.050 7.060
12 5.800 5.800 5.800 5.800 5.800 5.800 5.802
13 5.270 5.270 5.270 5.270 5.270 5.270 5.204
a
14 7.990 7.990 7.990 8.000 7.990 7.990 7.943
15 6.570 6.570 6.570 6.570 6.570 6.567 6.441
16 5.100 5.100 5.100 5.100 5.1005 5.100 5.259
17 8.000 8.000 8.000 8.000 8.000 8.000 8.000
a
18 7.134 7.134 7.1317 7.1368 7.134 7.134 7.136
19 6.070 6.070 6.070 6.070 6.070 6.053 6.041
20 5.820 5.820 5.820 5.820 5.8214 5.820 5.815
21 6.380 6.380 6.380 6.380 6.380 6.380 6.431
22 5.990 5.990 5.990 5.990 5.990 5.990 5.971
23 4.660 4.660 4.660 4.660 4.660 4.660 4.583
24 4.600 4.600 4.600 4.600 4.6001 4.600 4.600
25 3.580 3.580 3.580 3.580 3.5793 3.580 3.580
26 3.980 3.980 3.980 3.980 3.9788 3.980 3.965
a
27 3.900 3.8938 3.8961 3.9097 3.9009 3.903 3.902
28 3.980 3.980 3.980 3.980 3.9803 3.980 3.978
29 3.900 3.900 3.900 3.900 3.902 3.900 3.900
30 3.470 3.470 3.470 3.470 3.4702 3.470 3.479
31 3.200 3.200 3.200 3.200 3.2009 3.200 3.196
32 2.980 2.980 2.980 2.980 2.980 2.980 2.986
a
33 3.150 3.150 3.158 3.1687 3.1499 3.154 3.153
TAE C/T 1.43e13 121e13 369e13 0.0048 0.024 0.850
Test 0.0099 0.0245 0.0742 0.0065 0.007 0.151
Sum 0.0099 0.0245 0.0742 0.0113 0.031 1.001

Abbreviations: ANFIS, adaptive network fuzzy inference system; ANN, artificial neural network; DE, differential evolution; PSO, particle swarm optimization;
RF, resonant frequency; TAE, total absolute error; SM, surrogate model.
a
The bold values for the test; C/T: construction (C) for the SM or training (T) for the neuro-models.
10 of 12 USTUN ET AL.

TABLE 4 Comparative results of the measured and computed BW with the literature

BW (%)

This study Literature

Optimized Analytic Neuro-model Formula


4,5 19 20
Antenna no. Measured SM with DE SM with PSO SM ANN ANFIS 4
1 1.07 1.070 1.070 1.070 1.069 1.070 1.200
2 2.20 2.200 2.200 2.200 2.199 2.200 2.780
3 3.85 3.850 3.850 3.850 3.850 3.850 5.030
4 1.95 1.950 1.950 1.950 1.949 1.950 2.460
5 2.05 2.050 2.050 2.050 2.050 2.050 4.090
6 5.10 5.100 5.100 5.100 5.101 5.100 6.460
7a 4.90 4.899 4.882 4.860 4.560 4.926 6.170
8 6.80 6.800 6.800 6.800 6.800 6.800 8.120
9 5.70 5.700 5.700 5.700 5.699 5.700 7.120
a
10 7.70 7.600 7.600 7.603 7.811 7.716 9.160
11 10.90 10.900 10.900 10.900 10.899 10.900 11.720
12 9.30 9.300 9.300 9.300 9.299 9.300 10.420
13 10.00 10.000 10.000 10.000 10.001 10.000 11.150
a
14 16.00 16.000 16.000 16.002 15.954 16.059 15.160
15 13.60 13.600 13.600 13.600 13.601 13.600 13.140
16 15.90 15.900 15.900 15.900 15.899 15.900 15.110
17 17.50 17.500 17.500 17.500 17.499 17.500 17.000
a
18 18.20 18.197 18.136 18.205 18.345 18.304 17.770
19 17.90 17.900 17.900 17.900 17.877 17.900 17.340
20 18.00 18.000 18.000 18.000 18.023 18.000 17.470
21 19.00 19.000 19.000 19.000 19.004 19.000 18.420
22 20.00 20.000 20.000 20.000 20.000 20.000 19.290
23 18.70 18.700 18.700 18.700 18.699 18.700 18.010
24 20.90 20.900 20.900 20.900 20.919 20.900 21.260
25 20.00 20.000 20.000 20.000 20.000 20.000 19.660
26 20.60 20.600 20.600 20.600 20.600 20.600 20.850
a
27 20.30 20.289 20.253 20.601 20.237 20.307 20.610
28 20.90 20.900 20.900 20.900 20.873 20.900 21.400
29 21.96 21.960 21.960 21.960 21.948 21.960 22.260
30 21.50 21.500 21.500 21.500 21.545 21.500 21.910
31 21.60 21.600 21.600 21.600 21.571 21.600 21.730
32 20.40 20.400 20.400 20.400 20.405 20.400 16.070
a
33 21.20 21.217 21.171 21.345 21.265 21.183 21.550
TAE C/T 1.75e11 2.13e11 2.79e11 0.199 0 19.260
Test 0.132 0.328 0.589 0.770 0.229 4.660
Sum 0.132 0.328 0.589 0.969 0.229 23.920

Abbreviations: ANFIS, adaptive network fuzzy inference system; ANN, artificial neural network; BW, bandwidth; DE, differential evolution; PSO, particle
swarm optimization; TAE, total absolute error; SM, surrogate model.
a
The bold values for the test; C/T: construction (C) for the SM or training (T) for the neuro-models.
USTUN ET AL. 11 of 12

7 | C ON C L U S I ON

In this study, a computational model based on SM is proposed for the analysis of the RMAs with regard to RF and
BW. An SM based on RBF was, respectively, constructed and tested on the numbers of 27 and 6 RMAs, which were
obtained on a previously reported dataset of 33 measured RMAs. In general, the weighting vectors of SM have been
analytically derived through the input–output interrelationship in the construction. However, the obtained SM in this
way might not be tight enough with the interrelationship. A novel method based on evolutionary optimization, for
example, DE and PSO algorithms, was hereby proposed to determine the weighting vectors together with the factors of
power-exponential RBF. Even the fact that other well-known interpolation functions were essayed in the SM, the RBF
was the most successful. The SM constructed with and without DE and PSO was compared separately with the best
results of neuro-models and formulaic models in terms of the TAE reported elsewhere. The results show that the SM
optimized with DE computes, respectively, the RF and BW with the best TAE of 0.0099 GHz and 0.131%. The SM with
PSO predicts the RF and BW with TAE of 0.0245 GHz and 0.328%. On the other side, the best results of the neuro-
models regarding the TAE are 0.0113 GHz by ANN and 0.229% by ANFIS for the RF and BW. Meanwhile, the accuracy
of the SM is thus increased by 78% via optimizing the SM with DE. It is demonstrated that the proposed SM-based com-
putational scheme, which combines the precision of neuro-models and the simplicity of formulaic models, can be also
optimized by an evolutionary optimization algorithm. Although the SM so-called metamodel is simpler and handier
than the neuro-models, it is more accurate than the neuro-models. Therefore, a metamodel, which is easy, fast, and
handy is proposed to be a better candidate for CAD applications.

CONFLICT OF INTEREST
The authors declare no potential conflict of interests.

DATA AVAILABILITY STATEMENT


The dataset supporting the finding of this study is available in references 4–6.

ORCID
Abdurrahim Toktas https://orcid.org/0000-0002-7687-9061

R EF E RE N C E S
1. Şimşek S. An efficient neural network model for the input resonant resistance of coaxial fed rectangular microstrip antennas with thin
and thick substrates. Microwave Opt Technol Lett. 2006;48:2197-2199.
2. Lee KF, Luk KM, Kwai M. Microstrip Patch Antennas. Imperial College Press; 2011.
3. Bahl IJ, Bhartia P. Microstrip Antennas. Artech House; 1980.
4. Kara M. A novel technique to calculate the bandwidth of rectangular microstrip antenna elements with thick substrates. Microwave Opt
Technol Lett. 1996;12:60-64.
5. Kara M. A simple technique for the calculation of the bandwidth of rectangular microstrip antenna elements with various substrate
thicknesses. Microwave Opt Technol Lett. 1996;12:16-20.
6. Kara M. The resonant frequency of rectangular microstrip antenna elements with various substrate thicknesses. Microwave Opt Technol
Lett. 1996;11:55-59.
7. Kara M. Closed-form expressions for the resonant frequency of rectangular microstrip antenna elements with thick substrates. Micro-
wave Opt Technol Lett. 1996;12:131-136.
8. Akdagli A. A closed-form expression for the resonant frequency of rectangular microstrip antennas. Microwave Opt Technol Lett. 2007;
49:1848-1852.
9. Akdagli A. An empirical expression for the edge extension in calculating resonant frequency of rectangular microstrip antennas with
thin and thick substrates. J Electromagn Waves Appl. 2007;21:1247-1255.
10. Karaboga D, Güney K, Sagiroglu S, Erler M. Neural computation of resonant frequency of electrically thin and thick rectangular micro-
strip antennas. IEE Proc Microwaves Antennas Propag. 1999;146:155.
11. Sagiroglu S, Guney K, Erler M. Calculation of bandwidth for electrically thin and thick rectangular microstrip antennas with the use of
multilayered perceptrons. Int J RF Microwave Comput Aided Eng. 1999;9:277-286.
12. Guney K, Sagiroglu S, Erler M. Generalized neural method to determine resonant frequencies of various microstrip antennas. Int J RF
Microwave Comput Aided Eng. 2002;12:131-139.
13. Guney K, Gultekin SS. Artificial neural networks for resonant frequency calculation of rectangular microstrip antennas with thin and
thick substrates. Int J Infrared Millimeter Waves. 2004;25:1383-1399.
14. Sagiroglu S, Kalinli A. Determining resonant frequencies of various microstrip antennas within a single neural model trained using par-
allel Tabu search algorithm. Electromagnetics. 2005;25:551-565.
12 of 12 USTUN ET AL.

15. Guney K, Sarikaya N. Adaptive neuro-fuzzy inference system for computing the resonant frequency of electrically thin and thick rectan-
gular microstrip antennas. Int J Electron. 2007;94:833-844.
16. Merad L, Bendimerad FT, Meriah SM. Design and resonant frequency calculation of rectangular microstrip antennas. Int J Numer
Modell Electron Networks Devices Fields. 2011;24:144-153.
17. Chen F, Tian YB. Modeling resonant frequency of rectangular microstrip antenna using CUDA-based artificial neural network trained
by particle swarm optimization algorithm. Appl Comput Electromagn Soc J. 2014;29:1025-1034.
18. Guney K. A simple and accurate expression for the bandwidth of electrically thick rectangular microstrip antennas. Microwave Opt
Technol Lett. 2003;36:225-228.
19. Gultekin SS, Guney K, Sagiroglu S. Neural networks for the calculation of bandwidth of rectangular microstrip antennas. Appl Comput
Electromagn Soc J. 2003;18:46-56.
20. Guney K, Sarikaya N. Adaptive neuro-fuzzy inference system for the computation of the bandwidth of electrically thin and thick rectan-
gular microstrip antennas. Electr Eng. 2006;88:201-210.
21. Müller J. SOCEMO: surrogate optimization of computationally expensive multiobjective problems. INFORMS J Comput. 2017;29:
581-596.
22. Forrester AIJ, Sbester A, Keane AJ. Engineering Design via Surrogate Modelling. John Wiley & Sons, Ltd; 2008.
23. Koziel S, Ciaurri DE, Leifsson L. Surrogate-based methods. Stud Comput Intell. 2011;356:33-59.
24. Toktas A, Ustun D, Tekbas M. Multi-objective design of multi-layer radar absorber using surrogate-based optimization. IEEE Trans
Microwave Theory Tech. 2019;67:3318-3329.
25. Ustun D, Toktas A. Translational motion compensation for ISAR images through a multicriteria decision using surrogate-based optimi-
zation. IEEE Trans Geosci Remote Sens. 2020;58:4365-4374.
26. Carbas S, Toktas A. In: Ustun D, ed. Nature-Inspired Metaheuristic Algorithms for Engineering Optimization Applications. Springer
Singapore; 2021.
27. Toktas A, Ustun D, Erdogan N. Pioneer Pareto artificial bee colony algorithm for three-dimensional objective space optimization of
composite-based layered radar absorber. Appl Soft Comput J. 2020;96:1-12.
28. Toktas A, Ustun D. A triple-objective optimization scheme using butterfly-integrated ABC algorithm for design of multilayer RAM. IEEE
Trans Antennas Propag. 2020;68:5602-5612.
29. Dorigo M, Stützle T. Ant Colony Optimization. MIT Press; 2004.
30. Kennedy J, Eberhart R. Particle swarm optimization. Proceedings of ICNN'95 – International Conference on Neural Networks. Vol. 4.
IEEE, 1995:1942-1948.
31. Goldberg DE. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley Longman Publishing; 1989.
32. Toktas A, Kayabasi A, Sabanci K, Yigit E. Notch antenna analysis: an expression for calculation of the operating frequency. Microwave
Opt Technol Lett. 2017;59:1309-1313.
33. Mirjalili S, Lewis A. The whale optimization algorithm. Adv Eng Software. 2016;95:51-67.
34. Karaboga D, Basturk B. A powerful and efficient algorithm for numerical function optimization: artificial bee colony (ABC) algorithm.
J Global Optim. 2007;39:459-471.
35. Storn R, Price K. Differential evolution – a simple and efficient heuristic for global optimization over continuous spaces. J Global Optim.
1997;11:341-359.
36. MJD P. Radial basis functions for multivariable interpolation: a review. In: Mason JC, Cox MG, eds. Algorithms for Approximation.
Clarendon Press; 1987:143-167.
37. Box GEP, Wilson KB. On the experimental attainment of optimum conditions. J R Stat Soc Series B. 1951;13:1-45.
38. Krige DG. A statistical approach to some basic mine valuation problems on the Witwatersrand. J Chem Metall Soc South Mining Africa.
1951;52:119-139.
39. Müller J, Shoemaker CA. Influence of ensemble surrogate models and sampling strategy on the solution quality of algorithms for com-
putationally expensive black-box global optimization problems. J Global Optim. 2014;60:123-144.

How to cite this article: Ustun D, Toktas F, Toktas A. An optimized surrogate model using differential
evolution algorithm for computing parameters of antennas. Int J Numer Model. 2021;e2951. doi:10.1002/jnm.2951

You might also like