Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

2017 11th European Conference on Antennas and Propagation (EUCAP)

Accelerated Multi-Objective Design Optimization


of Antennas By Surrogate Modeling
and Domain Segmentation
Slawomir Koziel1,2, Adrian Bekasiewicz1,2, Qingsha S. Cheng3 and Song. Li3,4
1
Engineering Optimization & Modeling Center, Reykjavik University, 101 Reykjavik, Iceland
2
Faculty of Electronics, Telecommunications and Informatics, Gdansk University of Technology, 80-233 Gdansk, Poland
3
Department of Electrical and Electronic Engineering, Southern University of Science and Technology, Shenzhen, China
4
Department of Electronic System Engineering, University of Regina, Regina, Canada

Abstract—Multi-objective optimization yields indispensable (PSO) [8], [9], as well as multi-objective versions of modern
information about the best possible design trade-offs of an antenna metaheuristics (e.g., differential evolution [10] or firefly
structure, yet it is challenging if full-wave electromagnetic (EM) algorithm [11]). The most appealing feature of population-
analysis is utilized for performance evaluation. The latter is a
necessity for majority of contemporary antennas as it is the only based methods is their ability of generating the entire
way of achieving acceptable modeling accuracy. In this paper, a representation of the Pareto set in a single algorithm run. Their
procedure for accelerated multi-objective design of antennas is fundamental drawback is very high computational cost,
proposed that exploits fast data-driven surrogates constructed at typically thousands and tens of thousands of objective function
the level of coarse-discretization EM simulations, multi-objective evaluations. Clearly, such a cost might be prohibitive from the
evolutionary algorithm to yield an initial approximation of the
Pareto set, and response correction methods for design refinement point of antenna optimization, particularly if full-wave EM
(i.e., elevating the selected Pareto-optimal designs to the high- analysis is required for performance evaluation of the structure.
fidelity EM simulation model level). To reduce the computational A possible way of alleviating the aforementioned difficulty
cost of setting up the surrogate, the relevant part of the design is utilization of surrogate modeling techniques. A surrogate-
space (i.e., the part containing the Pareto front) is first identified assisted method for multi-objective antenna optimization has
through a series of single-objective optimization runs and
subsequently represented by a set of adjacent compartments with been proposed in [12] that exploits data-driven model,
separate surrogate models established within them. This implemented using kriging interpolation [13], and constructed
segmentation process dramatically reduces the number of training from sampled coarse-discretization EM simulation data of the
samples necessary to build an accurate model thus limiting the antenna at hand. The surrogate has been optimized using multi-
overall optimization cost. Our approach is demonstrated using a objective evolutionary algorithm (MOEA) to yield an initial
UWB monopole antenna and compared to a state-of-the-art
surrogate-assisted technique that does not use domain approximation of the Pareto set. The final Pareto set (at the
segmentation. level of high-fidelity EM antenna model) has been obtained by
means of output space mapping [12].
Index Terms—Antenna design, multi-objective optimization,
simulation-driven design, surrogate modeling, design space
The technique of [12] has been enhanced in [14] by
reduction, domain segmentation. constraining the search space to the interval spanned by the
extreme Pareto-optimal designs found by single-objective
optimization runs (one objective at a time) thus reducing the
I. INTRODUCTION
volume of the region that need to be sampled in the process of
Design of modern antennas is a complex process in which surrogate model construction. More involved design space
several (and often conflicting) objectives have to be taken into reduction methods were proposed in [15] and [16] that allow
account at the same time [1]. Due to stringent requirements for handling high-dimensional antenna design problems.
concerning reflection response, radiation figures, but also In this paper, we introduce a design space segmentation
antenna geometry (e.g., minimization of the footprint [2]), any approach which permits further reduction of the computational
practical design is a trade-off between the considered goals. cost of multi-objective antenna optimization as compared to
From the designer’s perspective, it is indispensable to identify [12]. The region containing the Pareto set is covered by a few
the best possible trade-offs between various figures of interest, intervals spanned by the Pareto-optimal designs identified by
often referred to as a Pareto front. This, however, can only be appropriate single-objective optimization runs. Separate data-
achieved through multi-objective optimization [3]. driven surrogates are constructed in each interval. Identification
A number of techniques have been developed for solving of the initial Pareto sets and design refinement are arranged as
multi-objective design problems, including those that adopt in [12] but they are done independently for each interval.
conventional numerical optimization techniques (e.g., weighted Computational benefits in terms of dramatic reduction of the
sum method [4] or goal attainment method [5]) yet the most number of necessary training data samples are a result of the
popular ones include population-based metaheuristics such as fact that the overall volume of the search space that needs to be
genetic algorithms (GAs) [6], [7], particle swarm optimizers sampled is considerably smaller than the volume of the initial

/17/$31.00 ©2017 IEEE 3265


2017 11th European Conference on Antennas and Propagation (EUCAP)

design space. Our methodology is demonstrated using an ultra- of the initial Pareto set, the corresponding high-fidelity-level
wideband (UWB) monopole antenna and favorably compared design xf(k) is obtained using the refinement procedure [12]
to the benchmark method of [12]. (1)
x (fk ) = arg min ( k ) F1 ( Rs ( x ) + [ R f ( xs( k ) ) − Rs ( xs( k ) )])
x , F2 ( x ) ≤ F2 ( xs )

II. MULTI-OBJECTIVE OPTIMIZATION USING SURROGATE FNobj ( x ) ≤ FNobj ( xs( k ) )
MODELING AND DESIGN SPACE SEGMENTATION
The refinement process attempts to improve the first objective
In this section, we formulate the proposed optimization without degrading the remaining ones. The term Rf(xs(k)) –
methodology. We start by recalling the surrogate-assisted Rs(xs(k)) in (1) “shifts” the surrogate model response to make it
algorithm [12]. Subsequently, we introduce the domain
perfectly aligned with the high-fidelity model at xs(k) (zero-
segmentation approach provide qualitative explanation of
order consistency [18]). This type of correction is known as
computational benefits that it may provide. Illustration
output space mapping [19]. Because the low- and high-fidelity
example, a UWB monopole antenna, is discussed in
Section III. EM models are normally well correlated, two iterations of (1)
are typically sufficient to find a refined design.
A. Multi-Objective Optimization Problem C. Design Space Reduction
We use Rf(x) to denote a response of an accurate EM- If the number of design parameters is small and their
simulation antenna model. A vector x represents design ranges are moderate, the kriging surrogate can be constructed
variables, normally. We consider Nobj design goals, Fk(Rf(x)), in the entire design space. However, for most practical cases
k = 1, …, Nobj. Examples of goals include minimization of the including high-dimensional ones (n ≥ 10 or even over 15), the
reflection coefficient over a frequency band of interest, initial design space has to be reduced in order to permit
reduction of the side lobes or reduction of the antenna surrogate model construction without using excessive number
footprint. If Nobj > 1, comparison of designs is realized using a of training samples.
so-called dominance relation [3]. For two arbitrary designs x We denote by x*(k), k = 1, …, Nobj, the extreme points of
and y, we say that y  x (y dominates over x) if Fk(Rf(y)) ≤ the Pareto front, i.e., the designs that are optimum with
Fk(Rf(x)) for all k = 1, …, Nobj, and Fk(Rf(y)) < Fk(Rf(x)) for at respect to the kth objective, i.e.,
least one k. x*(k) = argmin{x : Fk(Rcd(x))} (2)
The purpose of multi-objective optimization is to find a In vast majority of practical antenna design cases, the
representation of a Pareto set XP that consists of non- Pareto front will be contained in the interval spanned by these
designs [14]. Therefore, it is reasonable to restrict the multi-
dominated designs so that for any x ∈ XP, there is no y for
objective optimization process (consequently, surrogate model
which y  x [3]. It should be noted that the elements of the construction) to the region defined by the following reduced
Pareto set represent the best possible trade-offs between lower/upper bounds l/u as [14] (cf. Fig. 1(a))
conflicting objectives. In particular, improving the value of l = min{x*(1), x*(2), …, x*(Nobj)} (3)
one of the objectives of the Pareto-optimal design necessarily and
degrades at least one of the remaining objectives. u = max{x*(1), x*(2), …, x*(Nobj)} (4)
In practice, the space reduced this way is a tiny fraction of
B. Multi-Objective Optimization with Surrogate Models the initial one (volume-wise). This reduces the cost of
The basis of the optimization approach proposed here is a surrogate model construction (specifically, the number of
surrogate-assisted technique described in [12]. It allows for training data and the associated cost of EM simulations). As
handling an expensive EM-simulation model by exploiting a demonstrated in the literature, the method [12] in conjunction
fast surrogate Rs, which is a data-driven model (here, a with the above reduction scheme permits handling antenna
kriging interpolation one [10]) constructed from sampled design problems with up to 20 adjustable parameters. Still, the
coarse-discretization EM simulations (denoted as a low- number of training samples necessary to construct the
fidelity model Rcd). The initial Pareto set representation is surrogate may be considerable [12].
subsequently generated by optimizing Rs using a multi- D. Design Space Segmentation
objective evolutionary algorithm (MOEA) [3]. In [12], a
Further reduction of the number of training samples
standard implementation of MOEA is used with fitness
necessary to build the surrogate model can be achieved by
sharing, Pareto-ranking tournament selection, and mating
design space segmentation proposed in this work. Here, it is
restrictions [17]. formulated assuming two objectives, however, it can be
Because the initial approximation of the Pareto set is generalized for a larger value of Nobj.
obtained at the level of the coarse-discretization EM antenna Let d(1) = [d1(1) d2(1) … dn(1)]T = |x*(1) – x*(2)| be the size
model, an additional step is required to find Pareto-optimal vector of the design space X0 reduced as described in
designs at the high-fidelity EM modeling accuracy. This is
Section II.C. Its volume is V0 = ∏k = 1,…,ndk(1). Figure 1(a)
achieved by using response correction techniques. More
explains the utilized notation. For majority of practical
specifically, if xs(k), k = 1, …, K, is a set of selected elements

3266
2017 11th European Conference on Antennas and Propagation (EUCAP)

antenna design problems, the reduced space X0 contains the where the threshold F2.I(1) is set to place the intermediate point
entire Pareto front or majority of it [14]. Thus, the surrogate around the middle of the Pareto front, i.e., F2.I(1) = [F2(x*(1)) +
model for MOEA optimization can be created only in X0 F2(x*(2))]/2. The starting point for solving (5) is [x*(1) + x*(2)]/2
instead the entire design space. and it is already a good approximation of xI(1).
Design space segmentation is based on introducing In a general case of K intermediate points xI(l), l = 1, …, K,
intermediate points that allows us to further reduce the one would have K + 1 sub-domains XK.l, l = 1, …, K + 1 of the
volume of the design space regions that need to sample while volumes VK.l = ∏k = 1,…,ndk(I.l), where d(I.1) = |x*(1) – xI(1)|, d(I.l) =
still capturing the entire Pareto front. Figures 1(b) and 1(c) |xI(l–1) – xI(l)| for l = 2, …, K, and d(I.K+1) = |x*(2) – xI(K)|. The
illustrate this concept for one and two intermediate points, intermediate points are obtained by solving
respectively. In case of one point, xI(1), two reduced sub- xI(l ) = arg min ( l ) F1 ( Rcd ( x ) ) (6)
domains are created, X1.1 and X1.2. Their volumes are x , F2 ( x ) ≤ F2. I

V1.l = ∏k = 1,…,ndk(I.l), where d(I.1) = [d1(I.1) … dn(I.1)]T = |x*(1) – with F2.I(l) = (1 – αl)F2(x*(1)) + αlF2(x*(2)), αl = l/K, and the
xI(1)|, and d(I.2) = [d1(I.2) … dn(I.2)]T = |x*(2) – xI(1)|. The starting points being (1 – αl)x*(1) + αl x*(2).
intermediate point is obtained by solving In the most optimistic scenario, i.e., when K intermediate
xI(1) = arg min (1) F1 ( Rcd ( x ) ) (5) point split the range of parameters between x*(1) and x*(2) into
x , F2 ( x ) ≤ F2. I
K + 1 equal intervals, the total volume VK = VK.1 + … + VK.K+1
in relation to the volume V0 of the initial domain (with no
x*(1) F2 [F1(x*(1)),F2(x*(1))] segmentation) is given by
VK 1 (7)
=
V0 ( K + 1)n−1
For example, if K = 1 and n = 12 (twelve design
parameters), the ratio is 1/2048, whereas for K = 2 and n = 12,
we have VK/V0 ≈ 5⋅10–6. For n = 18, we have V1/V0 ≈ 7⋅10–6
X0 [F1(x*(2)),F2(x*(2))] and V2/V0 ≈ 7⋅10–9. Thus, the benefits of segmentation grow
x*(2) quickly both with n and K. Obviously, increasing K increases
F1
(a) the cost of obtaining the intermediate points due to solving (6)
so that using K larger than three of four is not recommended.
One should also remember that (7) describes the extreme
situation that does not occur in practice. The actual ratios are
larger due to the fact that intermediate points never split the
initial domain X0 into equal-size intervals. Still, as
demonstrated in Section III, design space segmentation leads
to considerable computational savings.
III. CASE STUDY: UWB MONOPOLE ANTENNA
A. Antenna Structure
(b) Consider a compact UWB monopole antenna shown in
Fig. 2 [20]. The structure is implemented on a 0.762 mm thick
Taconic RF-35 dielectric substrate (εr = 3.5, tanδ = 0.0018). It
consists of a trapezoid radiator with improved impedance
matching. Further bandwidth enhancement is obtained by
employing two slots within the radiator and one cylindrical
slit below the feed line. The vector of antenna design
parameters is x = [l0 l1 w1r w2 o2r o3r s1r s2r s4r s5r]T. Relative
variables are w1 = (0.5w2 – 0.5w0)w1r, o2 = 0.5w2o2r, o3 = (l1 –
s3)o3r, s1 = (0.5w2 – 0.5w0)s1r, s2 = l1s2r, s4 = (w2 – 2s5)s4r, and
s5 = 0.5(l0 – g)s5r. Parameters w0 = 1.7, o1 = 0.25, g = 0.5. The
(c)
Fig. 1. Design space segmentation: (a) initially reduced space X1 spanned by unit for all dimensions is mm.
the extreme Pareto-optimal designs x*(1) and x*(2); the surrogate model for The EM-simulation models of the structure are
MOEA optimization is established in entire X1; (b) design space segmentation implemented in CST Microwave studio and simulated using
with one intermediate Pareto-optimal point x*(3); two independent kriging its time domain solver. The high-fidelity model Rf of the
models are established in two subsets X2.k, k = 1,2; the overall volume of the
sampled region is considerably smaller than for the previous case, yet the structure consists of ~6,000,000 hexahedral mesh cells and its
entire Pareto set is covered; (c) design space segmentation with two average simulation time on a dual Xeon E5540 machine with
intermediate points: the overall volume of the search space subsets to be 6 GB of RAM is 30 min. The low-fidelity model Rcd is
sampled is further reduced. discretized using about 1,400,000 cells (simulation time: 5,5

3267
2017 11th European Conference on Antennas and Propagation (EUCAP)

min). Both EM models include SMA connector to ensure Figure 3 indicates that Pareto sets from initially reduced,
reliable antenna evaluation (see Fig. 2(b)). as well as two-, and three-fold segmented spaces are very
The following design objectives are considered: F1 – similar. Vertical discrepancy between the responses of
minimization of reflection in 3.1 GHz to 10.6 GHz bandwidth selected Rf model designs is below 0.5 dB. Moreover,
and F2 – reduction of the antenna footprint S(x) = w2(l0 + l1). similarity of the results is confirmed in Fig. 4.
The initial ranges of design variables for the considered The computational cost of the multi-objective optimization
structure are: l0 = [4 3 0 4 0 0 0.1 0 0.01 0.01]T and u0 = [24 in the initially reduced space corresponds to 342 Rf model
24 1 24 0.5 1 0.9 0.3 0.5 0.5]T. simulations (171 h of CPU-time). At the same time, the cost
of multi-objective design in two- and three-fold segmented
B. Numerical Results
spaces corresponds to 182 Rf (~91 h) and 112 Rf (~56 h)
The initially reduced design space is defined using the model simulations (including the overhead for determination
extreme Pareto-optimal designs obtained by sequential single- of intermediate designs), respectively (see also Table IV). In
objective optimizations (cf. Section II.C): x*(1) = [9.46 15.99 other words, the design space segmentation indeed allows for
0.38 15.39 0.06 0.08 0.69 0.09 0.15 0.26]T (optimal w.r.t. F1) significant reduction of the overall design cost without
and x*(2) = [9.13 13.11 0.26 12.38 0 0.12 0.74 0.19 0.05 0.43]T compromising the Pareto front quality.
(optimal w.r.t. F2).
For the purpose of this study, the acceptable RMS error of g
the kriging model Rs is set to 2.5%. Data acquisition/model o3
s5 s
construction process is iterative [2]. The search space is w2 o2
o1
4 w0
discretized using Latin Hypercube Sampling algorithm (10 s1
samples per iteration) [21]. Then, Rcd model is evaluated and s2
kriging model is constructed using obtained training samples. l1 l0
The generalization error is estimated using cross-validation (a) (b)
[13]. The process continues until desired accuracy of the Fig. 2. A compact monopole antenna [20]: (a) geometry with highlight on design
parameters, and (b) 3D visualization of the structure with SMA connector.
model is obtained.
For the initially reduced space, the desired accuracy of the TABLE I: PARETO DESIGNS OBTAINED IN INITIALLY REDUCED SPACE
Rs model is obtained for 1500 low-fidelity model training
Design Variables [mm]
samples. In the next step, the initial Pareto front F1 F2
l0 l1 w1r w2 o2r o3r s1r s2r s4r s5r
representation is determined by MOEA optimization of the xf(10) –12.5 379 9.33 15.67 0.31 15.18 0.03 0.11 0.71 0.10 0.06 0.34
surrogate (algorithm setup: 500 individuals; 100 iterations). xf(8) –12.1 364 9.36 15.40 0.28 14.72 0.04 0.11 0.71 0.12 0.06 0.35
Finally, the high-fidelity Pareto-optimal designs are found xf(6) –11.3 332 9.37 14.79 0.30 13.74 0.02 0.11 0.72 0.14 0.06 0.39
using the refinement technique of Section II.B. The final xf(3) –10.9 311 9.28 14.52 0.30 13.07 0.02 0.10 0.71 0.15 0.07 0.40
xf(1) –9.9 288 9.16 13.77 0.28 12.55 0.00 0.11 0.74 0.17 0.07 0.39
representation of Pareto front is shown in Fig. 3. Detailed
dimensions of the selected high-fidelity designs are given in TABLE II: PARETO DESIGNS OBTAINED IN TWO-FOLD SEGMENTED SPACE
Table I, whereas their corresponding frequency characteristics
Design Variables [mm]
are shown in Fig. 4(a). F1 F2
l0 l1 w1r w2 o2r o3r s1r s2r s4r s5r
The intermediate point for two-fold segmented search xf(10) –12.5 387 9.40 15.91 0.35 15.30 0.04 0.09 0.69 0.10 0.08 0.31
space is xI(1) = [9.37 14.79 0.3 13.78 0.04 0.11 0.7 0.15 0.07 xf(8) –11.8 368 9.42 15.49 0.36 14.76 0.06 0.09 0.69 0.13 0.10 0.32
0.36]T. Kriging models for X1.1 and X1.2 are constructed using xf(6) –11.0 315 9.24 14.56 0.30 13.22 0.02 0.12 0.70 0.16 0.05 0.42
550 and 150 samples, respectively. It should be noted that xf(3) –10.5 295 9.36 14.27 0.26 12.47 0.04 0.11 0.74 0.15 0.05 0.40
xf(1) –10.0 284 9.14 13.78 0.27 12.38 0.01 0.12 0.74 0.19 0.05 0.41
domain reduction w.r.t. initially reduced space is
V1/V0 = 1⋅10–3. Each surrogate is optimized using MOEA and TABLE III: PARETO DESIGNS OBTAINED IN THREE-FOLD SEGMENTED SPACE
refinement of the selected low-fidelity designs is performed.
Design Variables [mm]
The Pareto-front representation is shown in Fig. 3. F1 F2
l0 l1 w1r w2 o2r o3r s1r s2r s4r s5r
Dimensions of the selected high-fidelity designs are gathered xf(10) –12.6 390 9.45 15.91 0.38 15.37 0.05 0.08 0.69 0.09 0.12 0.28
in Table II, whereas their corresponding frequency responses xf(8) –12.1 365 9.44 15.47 0.36 14.64 0.03 0.08 0.66 0.10 0.09 0.31
are shown in Fig. 4(b). xf(6) –10.9 317 9.33 14.52 0.29 13.29 0.02 0.10 0.72 0.16 0.08 0.39
For the three-fold segmented space, the intermediate xf(3) –10.6 296 9.25 14.23 0.28 12.60 0.01 0.10 0.68 0.16 0.07 0.40
xf(1) –10.1 281 9.21 13.47 0.27 12.40 0.00 0.12 0.74 0.19 0.05 0.43
designs are xI(1) = [9.38 15.32 0.32 14.29 0.03 0.09 0.71 0.12
0.1 0.33]T and xI(2) = [9.29 14.5 0.29 13.2 0.02 0.11 0.72 0.16 TABLE IV: COST BREAKDOWN
0.08 0.41]T. The kriging interpolation models with desired
accuracy have been obtained using 160, 70, and 60 training Initially Two-fold Three-fold
reduced space segmentation segmentation
samples, respectively. Domain reduction is V2/V0 ≈ 3⋅10–5. Extreme and
Then, MOEA optimization of each model has been performed 200 Rcd 290 Rcd 320 Rcd
intermediate points
and the selected designs have been refined. The high-fidelity Data acquisition 1500 Rcd 700 Rcd 290 Rcd
representation of the Pareto set is shown in Fig 3. Dimensions MOEA optimization N/A N/A N/A
of the high-fidelity designs are given in Table III, whereas Refinement 30 Rf 30 Rf 30 Rf
their frequency characteristics are shown in Fig. 4(c). Total cost 342 Rf (171 h) 182 Rf (90.8 h) 112 Rf (56 h)

3268
2017 11th European Conference on Antennas and Propagation (EUCAP)

-9 ACKNOWLEDGEMENT
F1 (|S11| in-band) [dB]

The authors would like to thank Computer Simulation


-10
Technology AG, Darmstadt, Germany, for making CST
-11 Microwave Studio available. This work was supported in part
by the Icelandic Centre for Research (RANNIS) Grant
-12 152034051, and by National Science Centre of Poland Grant
2013/11/B/ST7/04325, and by the National Natural Science
-13 Foundation of China Grant 61471258.
275 300 325 350 375 400
2
F2 (Footprint) [mm ]
REFERENCES
Fig. 3. High-fidelity Pareto designs obtained in the initially reduced design
space (□), as well as search spaces with two- (○) and three-fold (∇) [1] X.-S. Yang, K.-T. Ng, S.H. Yeung, and K.F. Man, “Jumping genes
segmentation. multiobjective optimization scheme for planar monopole ultrawideband
antenna,” IEEE Trans. Antennas Prop., vol. 56, no. 12, pp. 3659-3666, 2008.
[2] S. Koziel, A. Bekasiewicz, and W. Zieniutycz, “Expedited EM-driven multi-
-5 objective antenna design in highly-dimensional parameter spaces,” IEEE
Antennas and Wireless Prop. Lett., vol. 13, pp. 631-634, 2014.
|S11| [dB]

-10 [3] K. Deb., “Multi-Objective Optimization Using Evolutionary Algorithms,”


New york: John Wiley & Sons, 2001.
-15 [4] V. Changkong and Y.Y. Haimes, Multiobjective Decision Making Theory
and Methodology, North-Holland, New York, USA, 1983.
[5] K.C. Tan, E.F. Khor, and T.H. Lee, Multiobjective evolutionary algorithms
-20
2 4 6 8 10 12 and applications, Springer-Verlag, 2005.
Frequency [GHz] [6] C.M. van Coevorden, S.G. Garcia, M.F. Pantoja, et.al., “Microstrip-patch
(a) array design using a multiobjective GA,” IEEE Ant. and Wireless Prop. Lett.,
vol. 4, pp. 100-103, 2005.
-5 [7] D. Ding and G. Wang, “Modified multiobjective evolutionary algorithm
based on decomposition for antenna design,” IEEE Trans. Antennas Prop.,
|S11| [dB]

-10 vol. 61, no. 10, pp. 5301-5307, Oct. 2013.


[8] N. Jin and Y. Rahmat-Samii, “Advances in particle swarm optimization for
-15 antenna designs: real-number, binary, single-objective and multiobjective
implementations,” IEEE Trans. Ant. Prop., vol. 55, no. 3, pp. 556-567, 2007.
-20 [9] S. Chamaani, S.A. Mirtaheri, and M.S. Abrishamian, “Improvement of time
2 4 6 8 10 12 and frequency domain performance of antipodal Vivaldi antenna using
Frequency [GHz] multi-objective particle swarm optimization,” IEEE Trans. Antennas Prop.,
(b) vol. 59, no. 5 pp. 1738-1742, May 2011.
[10] E. Mezura-Montes, M. Reyes-Sierra, and C.A. Coello Coello, “Multi-
-5 objective optimization using differential evolution: a survey of the state-of-
the-art,” in (Chakraborty, U.K. ed.) Advances in Differential Evolution, 173-
|S11| [dB]

-10 196, 2006.


[11] X.S. Yang, Engineering optimization: an introduction with metaheuristic
-15 applications. Hoboken: John Wiley & Sons, 2010.
[12] S. Koziel and S. Ogurtsov, “Multi-objective design of antennas using
-20 variable-fidelity simulations and surrogate models,” IEEE Trans. Antennas
2 4 6 8 10 12 Prop., vol. 61, no. 12, pp. 5931-5939, Dec. 2013.
Frequency [GHz] [13] T.W. Simpson, J. Peplinski, P.N. Koch, and J.K. Allen, “Metamodels for
(c) computer-based engineering design: survey and recommendations,”
Fig. 4. Frequency responses of the selected Pareto-optimal designs xf(10) (––), Engineering with Computers, vol.17, no. 2, pp. 129-150, July 2001.
xf(8) (– –), xf(6) (···), xf(3) (–·–), xf(1) (○○) obtained in: (a) initially reduced design [14] A. Bekasiewicz, S. Koziel, and W. Zieniutycz “Design space reduction for
space, as well as segmented space with (b) one intermediate point, and (c) expedited multi-objective design optimization of antennas in highly-
two intermediate points. Note that responses are visually similar. dimensional spaces,” to appear in S. Koziel, L. Leifsson, and X.S. Yang
(Eds.) Solving Computationally Extensive Engineering Problems: Methods
and Applications, Springer, 2014.
IV. CONCLUSION [15] S. Koziel and A. Bekasiewicz, “Fast multi-objective surrogate-assisted
design of multi-parameter antenna structures through rotational design space
In this work, a novel approach to computationally-efficient reduction,” IET Microwaves Antennas Prop., 2016.
multi-objective design optimization of antenna structures has [16] S. Koziel and A. Bekasiewicz, “Fast multi-objective optimization of narrow-
band antennas using RSA models and design space reduction,” IEEE
been presented. We exploit data-driven surrogate modeling Antennas and Wireless Prop. Lett., vol. 14, pp. 450-453, 2015.
and variable-fidelity EM simulations, as well as design space [17] C.M. Fonseca, Multiobjective genetic algorithms with application to control
segmentation. The segmentation process is crucial to engineering problems. PhD thesis, Department of Automatic Control and
Systems Engineering, University of Sheffield, UK, 1995.
considerably reduce the number of training samples required [18] N.M. Alexandrov and R.M. Lewis, “An overview of first-order model
for surrogate model construction, which leads to lowering the management for engineering optimization,” Optimization Eng., vol. 2, no. 4,
overall design optimization cost. Our approach has been pp. 413-430, Dec. 2001.
illustrated through the design of a UWB monopole antenna [19] S. Koziel, J.W. Bandler and K. Madsen, “Towards a rigorous formulation of
the space mapping technique for engineering design,” Proc. Int. Symp.
and compared with the reference surrogate-assisted Circuits, Syst., ISCAS, vol. 1, 2005, pp. 5605-5608.
optimization technique. As demonstrated, computational [20] S.K. Palaniswamy, Y. Panneer, M.G. Nabi Alsath, M. Kanagasabai, S.
benefits of the proposed approach are considerable: the Kingsly, and S. Subbaraj, “3D eight-port ultra-wideband (UWB) antenna
array for diversity applications,” IEEE Ant. Wireless Prop. Lett., 2016.
obtained savings are 47 percent for two-fold and 67 percent [21] B. Beachkofski and R. Grandhi, “Improved distributed hypercube
for three-fold segmentation. sampling,” American Institute of Aeronautics and Astronautics, paper AIAA
2002-1274, 2002.

3269

You might also like