A methodology for the multi-objective shape

optimization of very thin noise barriers

R. Toledo∗ , J. J. Aznárez, D. Greiner, O. Maeso

Instituto Universitario de Sistemas Inteligentes y Aplicaciones Numéricas en Ingenierı́a
(SIANI). Campus Universitario de Tafira. Universidad de Las Palmas de Gran Canaria.
Spain

Abstract

The approach of this paper is based on the evolutionary multi-objective opti-

mization (EMO) of very thin noise barrier models with improved performance,
whose topological designs were previously studied in [1] under the consideration
of single-objective optimization scenarios for idealized single-wire barrier mod-
els. To assume such a simplification of reality, the dual boundary element (BE)
formulation arises as the most appropriate strategy involving BE to avoid draw-
backs associated with the exclusive implementation of the standard BE formu-
lation. Based on an own implementation of the NSGA-II [2] algorithm, the 2D
analysis performed in this work focuses on the simultaneous optimization of two
objectives in conflict: the maximization of noise attenuation and the minimiza-
tion of the costs of all barrier material used, represented by the overall length of
the boundaries. Under this framework, two optimal sets of designs are obtained
for each model after multi-objective optimization processes: when considering a
random initial population and when including the best single-objective optimal
design. The results obtained show wide and uniformly spread-out Pareto fronts,
reflected in the geometric diversity featured by optimal designs.
Keywords: Noise barriers, Very thin bodies, Shape optimization,
Multi-objective Genetic Algorithms, Dual Boundary Element Formulation.

∗ Correspondingauthor
Email address: rtoledo@siani.es (R. Toledo∗ )

Preprint submitted to Computer Methods in Applied Mechanics and EngineeringMarch 16, 2016
 1. Introduction

In real-world applications, the existence of problems with just a desired

aim is a very rare scenario. Usually, multiple, often conflicting objectives arise
naturally in most optimization problems. The task of solving multi-objective
5 optimization problems is called multi-objective optimization (MO). Under this
framework, optimizing involves establishing a set of design variables which satis-
fies constraints and simultaneously optimizes the set of objective functions that
defines the desired aims of our problem. The issue is further complicated when
the intended goals are in conflict to each other, that is, when the fulfillment
10 of one objective function mitigates against the fulfillment of the other. There-
fore, using MO gives rise to a more comprehensive design scenario by providing
a set of trade-off optimal solutions (not a single one, as in the case of single-
objective optimization), popularly known as Pareto-optimal solutions. Due to
the multiplicity in solutions, these problems are proposed to be solved suitably
15 using evolutionary algorithms (EA’s), which use a population approach in the
search process. Generally speaking, any of the optimization methods available
may provide a satisfactory solution within the proposed framework of this pa-
per. However, EA’s are relatively new but very powerful techniques used to
find solutions to many real-world search and optimization problems. In this
20 line, the implementation of EA’s has been found to be a highly effective way
of finding multiple effective solutions in a single simulation run [3, 4] and has
largely proved its prominent role in successfully solving an important number
of science and engineering problems and applications [5–11].
The approach of this work is based on the evolutionary multi-objective opti-
25 mization (EMO) of very thin noise barrier models with improved performance,
whose designs were previously studied in [1] by means of single-objective op-
timization framework for idealized null-thickness barrier models. One of the
criterion involves maximization of the noise attenuation efficiency, while the re-
maining one deals with the minimization of the total length of the barrier, as
30 representative of the material used cost. It is easy to see that no one objective

2
 can be improved without degrading performance on the other. In other words,
in this case, reduction in length of barriers is associated, in the vast major-
ity of cases, with lower shielding efficiency. This work deals with optimization
problems performed by the combined use of a multi-objective genetic algorithm
35 (MOGA) and a code that implements a dual BE formulation for the assess-
ment of very thin road barriers considered as single-wire configurations. In this
respect, the dual approach arises as the most appropriate strategy involving
BE to address the proposed problems numerically by allowing us to assume a
simplification of reality by idealizing very thin elements as null-thickness type,
40 greatly facilitating the geometric definition of complex configurations (especially
when compared with the case of the faithful, detailed definition of real complex
volumetric designs) with no substantial influence on the acoustic performance
for the considered thickness of very thin sections [12]. In dealing with these
issues with BE often results in serious numerical drawbacks associated with the
45 integration of quasi-singular points if not to a singular system of equations, as
such geometric idealization makes every node of the discretization hold both
the pressure and the flux value at each side of it, i.e., 2n unknowns per n nodes
(see Figure 1). The MOGA software used in this study applies the NSGA-II [2]
algorithm, one of the references in the EMO field. In this sense, results here
50 presented are a step forward in the direction marked by previous works devel-
oped within the SIANI institute. As a brief background, in Maeso et al. [13]
the simultaneous minimization of both noise attenuation level differences with
respect a reference shielding curve and the effective height of Y-shaped barriers
was conducted. Later on, the multi-objective optimization of the shielding effi-
55 ciency and the cost of the barrier, represented by the overall boundary length,
was addressed in Greiner et al. [14], on the basis of Y-shaped barriers with upper
absorbing surface considering different receiver points’ schemes. A comparative
study between the strategy of introducing the optimal single-objective solution
in the initial multi-objective population (seeded approach) and the standard
60 approach was also presented. On the same line of this latter work, a single-
and multi-objective optimization of Y models with different surface treatments

3
 was carried out in [15]. Finally, in Deb et al. [16] the multi-objective problem
addressed in [13] was solved with increased performance by means of obtain-
ing automatic design learnings through the use of the so-called innovization
65 technique. All the aforementioned works addressed the issue implementing the
standard BE formulation for the sound attenuation prediction of Y-shaped vol-
umetric barriers.
Under this framework, two optimal sets of designs are obtained for each
model after multi-objective optimization processes: 1) when starting off with
70 a random initial population and 2) when including the best single-objective
optimal design (obtained in [1]) in the initial population. The comparison of the
optimal Pareto fronts derived from both processes, with equal number of fitness
function evaluations, are performed by means of the hypervolume indicator. In
both cases, two-dimensional sound propagation hypotheses are considered, i.e.,
75 an infinite, coherent mono-frequency source of sound and a noise barrier with no
geometric variation that stands on a flat plane (ground) of uniform admittance.
The problem is performed in the frequency domain with the usual assumptions
(Helmholtz equation): the medium (air) is modeled as homogeneous, elastic and
isotropic with no viscosity, under small disturbances and initially at rest with
80 no wind effects.
This paper is structured as follows: after this introduction, in section 2 the
modeling and discretization by implementing a dual BE formulation is described.
Following, in section 3 the outline of the bi-dimensional scenario under study,
the determination of the acoustic performance, the representation of barrier pro-
85 files and a description of the topological designs for the MO shape optimization
is provided. Section 4 deals with the multi-objective shape optimization frame-
work for the presented study. Finally, section 5 shows results and discussion
while section 6 covers the conclusions of the paper.

4
 Figure 1: (a) Generic thin barrier modeled as a volumetric structure. (b) Idealization of the
former barrier as a null boundary thickness profile.

2. Modeling and discretization by implementing a dual BE formula-

90 tion

The next lines are focused on the description of the implemented dual BE
formulation to make possible the numerical treatment of thin noise barriers
idealized as null boundary thickness profiles (see Figure 1). The special nature
of these types of barriers makes necessary the addition of a complementary
95 formulation (hyper-singular) that coupled with the conventional BE formulation
yields a compatible system of equations.

2.1. Singular boundary integral equation (SBIE). Classical BE formulation

The SBIE for the boundary point i to be solved by BE can be expressed as

follows:

∂G(k, r) ∂p
Z Z
ci p i + − p dΓ = G0 (k, r) + G(k, r) dΓ (1)
Γ ∂nj Γ ∂nj

100 This integral equality just involves the boundary of the barrier under in-
R
vestigation. The − symbol represents the integral along the boundary to be
understood in the Cauchy principal value sense, once the singularity around the
collocation point i has been extracted (ci ), as shown in Figure 2. In (1), p is
the acoustic pressure field over the barrier surface and G(k, r) is the half-space

5
105 fundamental solution (the acoustic pressure field when the source is placed at
the collocation point i over a plane with admittance βg - ground admittance
-) and ci is the free term. As a general rule: ci = θ/2π, where θ represents
the inner angle to the boundary measured in radians. It is easily shown that
ci = 0.5 for smooth boundaries.
110 The expressions of the fundamental solution and its derivative for a perfectly
reflecting ground (βg = 0) for bi-dimensional, harmonious problems are:

1
G(k, r) = [K0 (ikr) + K0 (ikr)] (2)
2π

 
∂G(k, r) ik ∂r ∂r
=− K1 (ikr) + K1 (ikr) (3)
∂nj 2π ∂nj ∂nj
being i the imaginary unit, k the wave number and r, r the distances to the
observation point from the collocation point and its symmetric point with re-
spect to the ground plane, respectively (see Figure 2). K0 and K1 are the Bessel
115 modified functions of order 0 and 1, respectively.
By applying (1) on a node i of the boundary discretization and interpolating
the variable with quadratic elements, the following can be written:

NE X
X 3 NE X
X 3
ci · p i + pkl · hik
l = p0 +
∗
qlk · glik (4)
k=1 l=1 k=1 l=1

with N E being the overall number of elements. The repeated application of (1)
on each node of the boundary discretization leads to the following system of
120 equations:

(Cs + H) · P = G · Q + G0 (5)

where Cs is a diagonal matrix whose entries involve the free term values at the
nodes of the discretization, P, Q are the pressure and the flux (the derivative of
the pressure with respect to the normal at each boundary node) vectors, G0 vec-
tor stores the values of the fundamental solution concerning the external noise
125 source and H, G are matrices whose entries are associated with the integration

6
 cores of the BEM formulation, involving just the variables of the problem along
the barrier boundary:

∂G(k, r)
Z Z
hik
l = φl dΓk ; glik = G(k, r) φl dΓk (6)
Γk ∂nk Γk

with i being the collocation point, k the element under integration and φl the
shape functions with quadratic approximation of the local variable ξ along the
130 element under integration, both to get the acoustic pressure level along the
boundary (7) and to fit the barrier profile (isoparametric elements). The sound
pressure in the element k can be then expressed as follows:

3
X
k
p = φl pkl = φ1 pk1 + φ2 pk2 + φ3 pk3 (7)
l=1

being:

1
φ1 (ξ) = ξ (ξ − 1)
2
φ2 (ξ) = 1 − ξ 2 (8)
1
φ3 (ξ) = ξ (ξ + 1)
2

2.2. Hyper-singular boundary integral equation (HBIE)

135 The HBIE for the boundary point i to be solved by BE can be written as
follows:

∂ 2 G(k, r)
 
∂pi ∂G(k, r) ∂p ∂G0 (k, r)
Z Z
ci += p dΓ = − dΓ + (9)
Γ∂ni
∂ni ∂nj Γ ∂ni ∂nj ∂ni
R R
where the = and − symbols represent the integral along the boundary to be
understood in the Hadamard finite part integral and in the Cauchy principal
value sense, respectively. As the Hölder condition [17] must be satisfied at the
140 collocation point i, the numerical treatment of the hyper-singular formulation
can be conducted either 1) by using discontinuous boundary elements or 2) by
moving the source towards inside the element in a non-nodal point (see Figure 4).
In both strategies and in all situations ci = 0.5 in (9).

7
 Figure 2: Source-image collocation after discretization of a M-shaped barrier with quadratic
elements (3 nodes). Integration over a generic element k. The collocation point is denoted by
i while the observation point in the element under integration is denoted by j (after element
discretization with Gauss point distribution).

Expressions (10) and (11) show the values of the derivatives of the funda-
145 mental solution implied in the hyper-singular integral equation (9):

 
∂G(k, r) ik ∂r ∂r
=− K1 (ikr) + K1 (ikr) (10)
∂ni 2π ∂ni ∂nI

∂ 2 G(k, r) (ik)2
 
∂r ∂r 1
= K2 (ikr) + K1 (ikr) ni · nj + (11)
∂ni ∂nj 2π ∂ni ∂nj r
 
∂r ∂r 1
K2 (ikr) + K1 (ikr) nI · nj
∂nI ∂nj r
As in (1), i is the imaginary unit, k the wave number and r, r the distances
to the observation point from the collocation point and its symmetric point with
respect to the ground plane, respectively. It is worth making a distinction here
regarding the normal vectors involved in the expressions above. nj is the normal
150 to the boundary at the integration point and ni (nix , niy ), nI (nix ,-niy ) represent

8
 the normal vectors to the real boundary at the collocation point (i) and at its
symmetric point (I) placed on a fictitious, symmetric boundary with respect
to the ground plane, respectively. K1 and K2 represent the Bessel modified
functions of order 1 and 2, respectively.
155 By applying (9) on a node i of the boundary discretization and interpolating
the variable with quadratic elements, the following can be written:

NE X
3 NE 3
X ∂G0 (k, r) X X k ik
ci · qi + pkl · mik
l = + ql · ll (12)
∂ni
k=1 l=1 k=1 l=1

with N E being the overall number of elements. The repeated application of (9)
on each node of the boundary discretization, as in the classical formulation,
leads to the following system of equations:

∂G0
M · P = L − Ch · Q +


(13)
∂ni
160 where Ch is a diagonal matrix with entry values of 0.5, P, Q are the pres-
sure and the flux (the derivative of the pressure with respect to normal at each
∂G0
boundary node) vectors, ∂ni array stores the values of the derivative of the
fundamental solution concerning the external noise source, and M, L are matri-
ces whose entries are associated with the integration cores of the hyper-singular
165 BEM formulation involving just the variables of the problem along the barrier
boundary:

∂ 2 G(k, r) ∂G(k, r)
Z Z
mik
l = φl dΓk ; llik = φl dΓk (14)
Γk ∂ni nj Γk ∂ni
with i being the collocation point, k the element under integration and φl the
shape functions with quadratic approximation, with ξ representing the local
coordinate within the element with side limits (-1, 1) (see Figure 4).
170 The numerical strategies employed in the evaluation of both the singular
and the hyper-singular BEM integrals have been developed and implemented in
a computer code by following the patterns proposed by Sáez et al. [18].

9
 2.3. Dual boundary element formulation

When dealing with very thin bodies numerical integration problems, involv-
175 ing quasi-singular points, may appear affecting to the barrier performance. The
idealization of such boundaries as non-thickness bodies not only solves the prob-
lem but also contributes to ease their geometric representation. With this aim,
the SBIE and the HBIE are applied separately ([12, 19]). Figure 3 facilitates
comprehension. The boundaries at both sides of the idealized null-thickness
180 bodies are represented by the discretization, with disparate values of acoustic
pressure and flux. The application of the classic formulation of the method,
based on the SBIE applied at both sides of null-width elements, yields a singu-
lar system of equations that does not allow the solution of the problem to be
obtained. However, the use of both the SBIE and the HBIE leads to a dual BE
185 formulation that offers a proper solution to address this issue.

Figure 3: (a) Geometric idealization as single-wire configuration of a barrier featuring a very

thin section. (b) Strategy used to avoid the singularity around the collocation point in the
dual approach for the treatment of elements idealized as null-thickness type (see e.g. [20, 21]).

Figure 3(a) represents an idealization of a generic thin body to be solved by

the dual BE formulation. After a discretization process, each node holds the
values of pressure and flux with respect to the boundary normal (p+ , q + , p− , q −
hereinafter). The strategy used to isolate the singularity of the method in this

10
190 type of domains can be seen in Figure 3(b) [20, 21]. Thus, the BE expression
for these boundaries can be written as follows:

N N

X
X
0.5 p+ + +
G+ +


i + p −
i + H p
j j + H p
− −
j j = j qj + Gj qj
− −
+ G0 (k, r) (15)
j=1 j=1

being N the overall nodes number of the discretization over the boundary. Tak-
ing into account that n+ = −n− , it is easily shown that:

Hj+ = −Hj− ; G+ −
j = Gj (16)

Considering the Robin boundary condition:

qj = −i k βΓ pj (17)

195 the following matrix expression is obtained:

(0.5 I∗ + H + i k β G) · P = G0 (18)

being I∗ a matrix with the following form:

 
1 1 0 0 0 0 ··· 0 0
 
0 0 1 1 0 0 · · · 0 0
 
I∗ = 
 .. .. .. .. .. .. .. .. .. 
 (19)
. . . . . . . . .
 
0 0 0 0 0 0 ··· 1 1
that allows us to consider the contribution of the free term at both sides of the
discretization nodes.
As previously mentioned, the application of the SBIE at each collocation
200 point of the discretization (nodes) yields an incompatible system of equations
(two unknowns exist for each node), making the problem intractable. However,
the application of the HBIE as a complementary formulation permits us to
address the subject successfully by introducing the remaining equations that

11
 allows the problem to be solved. The hyper-singular expression concerning
205 these types of geometries is then obtained:

N N
∂p+
 
∂p− X
X
i i
Mj+ p+ L+ +
− − − −


0.5 + + + j + M p
j j = j qj + Lj qj (20)
∂ni ∂n+
i j=1 j=1

where:

∂p−
i
= −qi ; Mj+ = −Mj− ; L+
j = Lj
−
(21)
∂n+
i
Considering the Robin boundary condition (17), the following matrix ex-
pression can be written:

∂G0
[i k β (0.5 I∗ + L) + M] · P = (22)
∂ni
Finally, gathering expressions (18) and (22) the following matritial system
210 is obtained for the proper assessment of noise barrier configurations idealized as
bodies with null section:
 
0.5 I + H + i k β G
∗    


 
 






 G0 

     
 · P = (23)
    
 ∂G0 
     

  
 
  
i k β (0.5 I∗ + L) + M ∂ni

Once the variables on the barrier boundary are known, the acoustic pressure
values at any internal point (receiver position) can be easily obtained, as usual,
by applying the standard BE formulation:

N
X
pi = G0 (k, r) − (hj + i k β gj ) pj (24)
j=1

215 2.4. Discretization criterion applied

At this point it is important to note that the discretization process is frequency-

dependent. In this regard, all simulations in this work have been performed with
barrier boundary discretization featuring four elements per wavelength at least.

12
 A good convergence of results by applying this meshing criterion has been ob-
220 served in previous numerical experiments carried out by the researcher of this
work. Furthermore, this criterion is in line with other works [22, 23].
Some special procedures to tackle both sharp-angled boundaries and sharp
angles between boundaries are considered. In order to assure the convergence
of the numerical integrations of nearly singular integrals, the computing code
225 used in this work implements two strategies. One of them is based on the pro-
cedure proposed by Telles [24], consisting in the reallocation and concentration
of the Gauss points around the point with the minimum r distance within the
element under integration. The other strategy consists in the subdivision of the
associated element from the barrier discretization into multiple sub-intervals,
230 depending on the minimum r distance to the quasi-singular point. The final re-
sult is the overall sum of the numerical integration applied to each sub-interval
of the element.

2.5. Collocation point strategy performed

As for the SBIE formulation, the code implemented makes use of nodal
235 collocation with the exception of the nodes placed at non-connected extremes
of boundaries, where a non-nodal collocation strategy is employed.
Figure 4 represents the strategy used in the implementation of the hyper-
singular BEM formulation when assessing the acoustic efficiency of the road
barriers presented in this document. As previously mentioned, a non-nodal
240 collocation is required in the extreme nodes of the elements. Extensive references
concerning the choice of δ value can be found in scientific literature (e.g., [23, 25])
for discontinuous elements.
For the cases here presented, the collocation point strategy has been carried
out with a well-proven distance of δ = 5% of the element length for the point
245 displacement towards inside. The election of this value has been substantiated
by previous numerical experiments carried out by the authors of this work. Even
so, differences in results associated with the election of δ (within the framework
of the ranges used in the bibliography) are not significant, particularly given the

13
 Figure 4: Non nodal collocation points at the bound limits of the element when dealing with
the hyper-singular BEM formulation: (1) collocation point PC1; (2) collocation point PC2;
(3) collocation point PC3.

fact that sound pressure levels of interest here are those neither at the barrier
250 boundaries nor at the barrier corners but at the receiver points.

3. Problem definition

3.1. Bi-dimensional configuration

Figure 5 represents the general configuration of the study. It deals with a

source of sound placed on a ground with a perfectly reflecting surface (βg = 0)
255 at ds = 9.5 m from the feasible region, parallel to an infinite thin cross-section
noise barrier. A trapezoidal section holds the area for feasible profiles, defined
by the limited barrier projection to the ground, that is dp = 1.0 m, and the
maximum effective height to be achieved, that is heff = 3.0 m at the median of
the rectangle trapezium.
260 A grid of 4x4 receivers is considered. The first line of receivers lays on
the ground and the remaining ones are placed at different heights, vertically
separated by ∆y = 1.0 m. A horizontal distance of ∆x = 2.0 m among them is
considered. The nearest receivers to the side limit of the feasible region are

14
 Figure 5: Bi-dimensional configuration for the MO shape optimization of noise barriers with
improved performance. Dimensions expressed in meters.

dr = 2.0 m away. The proximity of the receivers to the barrier is motivated by

265 the fact that the barrier performance in near regions is more affected by the
shape design rather than by the effective height, as occurs in non-near regions.

3.2. Determination of the acoustic efficiency

The determination of the shielding efficiency of a barrier is well defined by

means of the insertion loss coefficient (IL) that in the harmonic problem, for
270 every frequency from the analyzed noise source, is defined as usual:

 
PB
IL = −20 · log10 [dB] (25)
PHS
on every frequency of the band spectrum. As can be seen, this coefficient repre-
sents the difference of sound pressure levels at the receiver points in the situation
with (PB ) and without (PHS ) considering the barrier.
With the purpose of conducting an optimization process where the excitation
275 is represented by a noise source pulsing at every frequency of the band spectrum,

15
 the efficiency of the barrier for a specific receiver can be written in terms of the
broadband insertion loss (ILtotal ):
 
NF
X
(Ai −ILi )/10 
 10
 
 i=1
ILtotal = −10 · log10  NF  [dB(A)] (26)

 X 
 10Ai /10 
i=1

being NF the studied spectrum number of frequencies (in the analyses conducted
here, NF = 14), Ai the spectrum A-weighted noise level and ILi the insertion
280 loss value for sources pulsing at every frequency of the spectrum, according
to (25).
In all the studies performed in this paper, the noise source is characterized by
using the UNE-EN 1793 [26] normalized traffic noise spectrum for third-octave
band center frequencies, ranging from 100 to 2.000 Hz, the same used by the
285 Spanish Technical Building Code (CTE) [27].

3.3. Barrier definition

Following [15, 28], a simple procedure to mathematically represent the geom-

etry of barriers is proposed. The design points of the screen model are defined in
a systematic, simple way in a reference domain as a previous step to the barrier
290 profile generation in the real space. In short, the transformed domain holds the
set of design variables of the model under study, denoted by (ξi , ηi ), and repre-
sents the rectangular search space for the GA (see left part of Figure 6). Every
(ξi , ηi ) point in the transformed domain has its image (xi , yi ) in the Cartesian
space, that is the real domain where the barrier operates.
295 The transformation of Figure 6 can be expressed as follows:
         
xi   xm   xm   xm   xm 
1 2 3 4
= γ1 + γ2 + γ3 + γ4 (27)
yi  ym  ym  ym  ym 
1 2 3 4

where:

16
 Figure 6: Bi-dimensional coordinate systems. Dimensions expressed in [m].

1   1  
γ1 = −ξ 1 − η ; γ2 = +ξ 1−η
2 2 (28)
1  1 
γ3 = η + ξ ; γ4 = η −ξ
2 2
and:

dp dp
xm m
1 = x4 = − ; xm m
2 = x3 =
2 2
    (29)
ds + dp ds
y1m = y2m = 0 ; y3m = heff ; y4m = heff
ds + (dp /2) ds + (dp /2)
In this paper heff = 3.0 m is proposed. This value and the maximum barrier
projection to the ground dp have been chosen according to the geometric dimen-
300 sions of the barriers studied herein and present in the bibliography. Both latter
parameters define the feasible region by generating a trapezoidal search space
in the Cartesian barrier domain (see right part of Figure 6). Its final dimensions
are dependent, logically, of the placement of the noise source (ds ).

3.4. Topological designs for MO optimization

305 This study intends to outline the most common and effective geometric pat-
terns, acoustically speaking, present in noise barriers. To do this and to ex-

17
 pose the scope and features of the method, eight different topological designs
are studied (see Figure 7), based either on an overall shape or in a top edge
optimization of the barrier model, depending on the case, featuring perfectly
310 reflecting surfaces (βb = 0).
As for the overall shape design optimization, five barrier models are pro-
posed: models a) and b) (polygonal-shaped designs), models c) and d) (splines-
based models) and model e) (Y-shaped barrier) (see left part of Figure 7). As for
the optimization of the top edge of the barrier, model f) (tree-shaped barrier),
315 model g) (Y-variant model) and model h) (fork-shaped model) are studied
(right part of Figure 7). Such models are based on a set of points defined by
design variables in a transformed domain proposed by the evolutionary algo-
rithm, according to the geometric model definition. It is worth stressing that
contrary to the general approach (Figure 6), the feasible space is constrained
320 to the barrier top in the Cartesian domain for models f), g) and h). A proper
definition and further details of these models can be found in [1].

4. Multi-objective optimization framework

The study here presented deals with optimization problems performed by the
combined use of a multi-objective genetic algorithm (MOGA) and a code that
325 implements a dual BE formulation for the assessment of road barriers featuring
very thin sections with perfectly reflecting surfaces. The MOGA software used
in this study applies the NSGA-II [2] algorithm, one of the references in the EMO
field. This code performs a uniform crossover operator applied on a population
size of 100 individuals. The considered mutation rate is 1/nch, where nch is the
330 chromosome length (nch = 8xn, with n being the overall number of the design
variables - of 8 bits precision each -). Codification of such design variables was
carried out with the Gray Code.

18
Figure 7: Proposed models for the MO shape optimization of noise barriers with improved
performance.

19
 Five MO independent optimization runs of each single-wire topological model
(as representative of their corresponding very thin profile) are conducted. To
335 ensure a fair comparison among the optimal solutions of a particular model
both after including the best single-objective solution (Case B ) of such model
(obtained and presented in previous work [1]) in the initial population and after
considering random initial individuals (Case R), different stop criteria are con-
sidered for each MO case. According to this, the stop condition is met for 400
340 and 600 generations of the MO optimization process for Case B and Case R,
respectively. Recalling that the best optimal barriers after single-objective op-
timizations were obtained for 20.000 objective functions evaluations, both cases
performed the same number of assessments of the objective functions (60.000).

4.1. Definition of the conflicting objectives

345 One of the aims of the multi-objective optimization presented in this work is
the search of barrier designs with ever-increasing screening performance (maxi-
mization of the shielding efficiency). This criterion can be properly determined,
as is well known, by the broadband insertion loss. Its representation in terms
of the objective function can be expressed as follows:

NR
X
OF1 = ILtotalj /NR [dB(A)] (30)
j=1

350 being ILtotalj the broadband IL (see (26)) for each receiver and NR the total
number of receivers considered in the study.
Additionally, the minimization of the costs of the material used for the bar-
rier profile is intended. This second objective can be expressed in terms of the
sum of the boundary lengths (see Figure 8), as follows:

NB
X NB q
X
OF2 = Lk = (xfk − xik )2 + (ykf − yki )2 [m] (31)
k=1 k=1

355 with Lk being the length of each boundary and NB the overall number of bound-
aries of the considered barrier.

20
 Figure 8: Total length of the barrier as a factor to be minimized along the multi-objective
optimization. Example of a Y-shaped barrier model.

It is not difficult to understand at this stage that both criteria are in conflict:
reducing the barrier length affects adversely, in the large majority of cases, the
acoustic efficiency.

360 5. Results and discussion

As stated above, the models to be optimized in this study were presented and
described in detail in an authors’ previous work [1] under a single-objective op-
timization scenario. Unlike this latter, the analysis performed here just consider
perfectly reflecting barrier surfaces.
365 Following, results concerning multi-objective optimization of such models
are shown.

5.1. Designs after MO optimization

Figures 9-16 collect the results after the multi-objective optimizations of each
model. The upper graphs show the accumulated non-dominated front solutions
370 of the five runs for both after starting off with a random initial population (Case

21
 R) and when including the best single-objective solution in the initial population
(Case B ). Some of these solutions are labeled from 1 to 7 in both cases and
the corresponding geometric representation of these solutions are shown at the
bottom of such graphs associated with their objective function values in brackets
375 (OF1 , OF2 ).
In order to determine whether the inclusion of the best single-objective so-
lutions (Case B ) is a real asset in the improvement of the models studied, the
comparison between the optimal Pareto front of this case and that from Case R
is conducted by applying the hypervolume measure (HV). Data concerning this
380 value for each optimization run are collected in Table 1 for both cases. Addition-
ally, some information from this table is represented graphically in Figures 17
and 18 by vertical lines, whose extremes correspond to the lowest and the high-
est HV value found along the optimization runs of each model, while the dot
spot represents the mean value. However, data interpretation may not be read-
385 ily apparent, so statistical analysis is required to draw clear inferences from it.
The sets of HV measures from both cases shown in Table 1 are checked for
normality (i.e., if the data set can be reasonably modeled as a Gaussian distri-
bution) by applying the Lilliefors test [29] with a significance level of 5%. In
case the null hypothesis (the statement under test) is accepted in both cases, a
390 two-tailed t-Student test [30] for unequal variances is then applied to prove if
there is statistical significant difference between the samples means at the 95%;
on the contrary, if the null hypothesis is rejected for just one of the cases under
analysis, the Wilcoxon signed rank test [31] is applied. In these latter cases,
the rejection of the the null hypothesis allows the direct comparison between
395 the mean HV measure of Case B and Case R, while its acceptance implies that
distributions are not different (no comparison is applicable).

22
Table 1: Hypervolume measure (HV), mean value (HV) and standard deviation (σ) of the
optimal Pareto fronts of barrier models when considering a random initial population (R)
and when including the best single-objective solution (B ) in the optimization runs. Reference
point: (OF1 , OF2 ) = (0, 20).

Run 1 Run 2 Run 3 Run 4 Run 5 HV σ

R 316.63 315.96 316.34 316.27 316.13 316.27 0.25

a) 3-sided polygonal
B 324.77 325.03 324.85 324.49 324.62 324.75 0.21

R 316.41 321.63 330.68 315.88 315.63 320.05 6.44

b) 5-sided polygonal
B 346.44 346.22 346.53 346.53 345.88 346.32 0.28

R 332.42 331.47 291.56 329.35 322.54 321.47 17.16

c) 3-cubic splines
B 325.14 329.41 330.15 324.66 324.75 326.82 2.72

R 312.09 325.40 326.31 318.33 330.27 322.48 7.23

d) 5-cubic splines
B 327.86 328.83 325.53 330.22 328.21 328.13 1.71

R 327.97 324.20 334.71 332.93 324.42 328.85 4.82

e) Y-shaped
B 335.89 335.87 335.38 335.80 335.88 335.76 0.22

R 338.20 338.53 338.72 332.40 337.41 337.05 2.65

f) Tree-shaped
B 339.08 339.12 338.85 338.60 339.00 338.93 0.21

R 325.62 328.87 326.89 339.64 343.80 332.96 8.21

g) Y-variant-shaped
B 337.04 336.24 336.90 333.89 338.46 336.51 1.67

R 325.81 319.80 325.44 325.09 325.51 324.33 2.55

h) Fork-shaped
B 332.50 332.27 332.24 332.43 332.52 332.39 0.13

23
 9.0
1
Case B: best single-objective solution in the initial population
8.5 Case R: random initial population
Reference straight barrier
8.0

7.5

7.0
2
6.5
OF2 [m]

6.0 3
5.5

5.0

4.5 1
4.0 4
2
3 5
3.5
4 7
3.0
5 6 6 7
2.5
-20.0 -19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5 -16.0 -15.5 -15.0 -14.5 -14.0 -13.5
OF1 [dB(A)]
1(-19.31, 8.70) 2(-18.83, 6.63) 3(-18.64, 5.77) 4(-17.75, 3.77) 5(-17.31, 3.36) 6(-15.08, 2.91) 7(-14.02, 2.85)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-18.58, 4.19) 2(-18.29, 4.04) 3(-17.93, 3.85) 4(-17.71, 3.37) 5(-15.75, 3.00) 6(-14.59, 2.87) 7(-14.01, 2.85)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 9: Model a): 3-sided polygonal shaped barrier results. Non-dominated solutions for
Case B and Case R.

24
 5.5
Case B: best single-objective solution in the initial population
Case R: random initial population
1 Reference straight barrier
5.0

1 2

4.5 3
OF2 [m]

2 3
4.0 4

3.5 4

5
6
3.0 7
5 6
7

2.5
-21.0 -20.5 -20.0 -19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5 -16.0 -15.5 -15.0 -14.5 -14.0 -13.5
OF1 [dB(A)]
1(-20.55, 4.63) 2(-19.33, 4.17) 3(-18.78, 4.01) 4(-18.01, 3.67) 5(-16.16, 3.11) 6(-15.56, 3.01) 7(-14.00, 2.90)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-19.49, 5.03) 2(-19.32, 4.66) 3(-18.79, 4.46) 4(-18.00, 3.96) 5(-16.16, 3.09) 6(-15.81, 3.01) 7(-14.01, 2.85)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 10: Model b): 5-sided polygonal-shaped barrier results. Non-dominated solutions for
Case B and Case R.

25
 4.5
Case B: best single-objective solution in the initial population
Case R: random initial population
1 Reference straight barrier
1
2
4.0
OF2 [m]

3.5
2 3

4 5
3.0
5 6
7
4 7
6

2.5
-19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5 -16.0 -15.5 -15.0 -14.5 -14.0
OF1 [dB(A)]
1(-19.35, 4.17) 2(-19.26, 3.53) 3(-18.18, 3.37) 4(-16.71, 2.92) 5(-15.36, 2.91) 6(-14.68, 2.87) 7(-14.05, 2.85)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-19.49, 4.11) 2(-19.24, 4.03) 3(-18.18, 3.30) 4(-16.72, 3.06) 5(-16.20, 3.01) 6(-14.68, 2.86) 7(-14.02, 2.85)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 11: Model c): 3-cubic splines-shaped barrier results. Non-dominated solutions for Case
B and Case R.

26
 8.0
1 Case B: best single-objective solution in the initial population
7.5 Case R: random initial population
Reference straight barrier
7.0

6.5

6.0
OF2 [m]

5.5

5.0

4.5
1 2
4.0
2 3
3.5 4
3
4 5
7
3.0
7
5 6 6
2.5
-19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5 -16.0 -15.5 -15.0 -14.5 -14.0
OF1 [dB(A)]
1(-19.40, 7.66) 2(-19.25, 4.17) 3(-19.11, 3.66) 4(-17.13, 2.98) 5(-15.85, 2.97) 6(-15.35, 2.88) 7(-14.31, 2.86)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-19.36, 4.11) 2(-19.26, 3.98) 3(-19.12, 3.43) 4(-17.14, 2.94) 5(-16.20, 2.88) 6(-15.05, 2.87) 7(-14.26, 2.87)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 12: Model d): 5-cubic splines-shaped barrier results. Non-dominated solutions for
Case B and Case R.

27
 6.0
1 1 Case B: best single-objective solution in the initial population
Case R: random initial population
5.5 2 Reference straight barrier

5.0

2
4.5
3
OF2 [m]

4.0 3

4
3.5
4
5
3.0

5 6
2.5
7 7
6
2.0
-20.0 -19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5 -16.0 -15.5 -15.0 -14.5 -14.0 -13.5 -13.0 -12.5 -12.0 -11.5 -11.0 -10.5 -10.0
OF1 [dB(A)]
1(-19.29, 5.88) 2(-18.83, 4.45) 3(-18.51, 3.81) 4(-17.41, 3.45) 5(-15.70, 3.02) 6(-13.28, 2.61) 7(-10.55, 2.18)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-19.22, 5.92) 2(-18.81, 5.67) 3(-18.53, 4.23) 4(-17.43, 3.43) 5(-14.60, 2.80) 6(-11.52, 2.24) 7(-10.05, 2.16)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 13: Model e): Y-shaped barrier results. Non-dominated solutions for Case B and Case
R.

28
 5.0
Case B: best single-objective solution in the initial population
Case R: random initial population
11

2
4.5 2

3
OF2 [m]

4.0
4

4 5
3.5 5
6
6
7 7

3.0
-21.0 -20.5 -20.0 -19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5 -16.0 -15.5 -15.0 -14.5
OF1 [dB(A)]
1(-20.53, 4.75) 2(-19.96, 4.58) 3(-18.87, 4.18) 4(-17.25, 3.81) 5(-16.78, 3.57) 6(-16.17, 3.44) 7(-15.23, 3.31)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-20.50, 4.74) 2(-19.97, 4.56) 3(-18.87, 4.33) 4(-17.26, 3.67) 5(-16.78, 3.71) 6(-15.70, 3.37) 7(-14.98, 3.31)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 14: Model f): Tree-shaped barrier results. Non-dominated solutions for Case B and
Case R.

29
 5.0
Case B: best single-objective solution in the initial population
Case R: random initial population

1
2

1 2
3

3
OF2 [m]

4.5 4

4 5

5 6
7
7
6

4.0
-22.0 -21.5 -21.0 -20.5 -20.0 -19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5
OF1 [dB(A)]
1(-21.51, 4.81) 2(-21.00, 4.76) 3(-20.96, 4.63) 4(-20.20, 4.45) 5(-19.13, 4.33) 6(-18.20, 4.27) 7(-17.43, 4.23)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-21.12, 4.68) 2(-20.99, 4.66) 3(-20.92, 4.58) 4(-20.20, 4.41) 5(-19.16, 4.31) 6(-18.21, 4.24) 7(-16.80, 4.21)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 15: Model g): Y-variant-shaped barrier results. Non-dominated solutions for Case B
and Case R.

30
 6.0
Case B: best single-objective solution in the initial population
1 Case R: random initial population
1

2
5.5 2
3

3 4
OF2 [m]

5.0 4

5
5
4.5
6
6
7
7
4.0
-21.5 -21.0 -20.5 -20.0 -19.5 -19.0 -18.5 -18.0 -17.5 -17.0 -16.5 -16.0 -15.5 -15.0
OF1 [dB(A)]
1(-21.27, 5.75) 2(-21.05, 5.47) 3(-20.59, 5.39) 4(-20.15, 5.18) 5(-17.96, 4.63) 6(-15.91, 4.26) 7(-15.38, 4.20)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
1(-20.80, 5.78) 2(-20.79, 5.63) 3(-20.61, 5.33) 4(-20.17, 5.06) 5(-17.60, 4.55) 6(-16.73, 4.37) 7(-15.38, 4.20)
3.5

3.0

2.5

2.0
y [m]

1.5

1.0

0.5

0.0
-0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5 -0.5 0.0 0.5
x [m] x [m] x [m] x [m] x [m] x [m] x [m]

Figure 16: Model h): Fork-shaped barrier results. Non-dominated solutions for Case B and
Case R.

31
 330 350
Model a). Random initial population Model b). Random initial population
Model a). Initial population with Model b). Initial population with
best single-objective solution best single-objective solution

345

340
325
Hypervolume measure (HV)

335

330

320
325

320

315 315
335 335
Model c). Random initial population Model d). Random initial population
Model c). Initial population with Model d). Initial population with
best single-objective solution best single-objective solution
330

330
325

320
Hypervolume measure (HV)

325
315

310
320

305

300
315

295

290 310

Figure 17: Hypervolume measures of the optimization runs for Case B and Case R. Models
a) to d).

32
 340 340
Model e). Random initial population Model f). Random initial population
Model e). Initial population with Model f). Initial population with
best single-objective solution best single-objective solution

335
Hypervolume measure (HV)

330 335

325

320
350
Model g). Random initial population
Model g). Initial population with
best single-objective solution
330
335
Model h). Random initial population
Model h). Initial population with
best single-objective solution

345
330
Hypervolume measure (HV)

340

325

335

320
330

325 315

Figure 18: Hypervolume measures of the optimization runs for Case B and Case R. Models
e) to h).

5.2. Discussion

According to the disclosed results, the following analysis is drawn:

• In all models, the value of the average hypervolume measures is always

33
400 better (higher) in Case B than in Case R.

• In all models, the value of the standard deviation of the hypervolume

measures is always better (lower) in Case B than in Case R.

• In four out of the eight models studied, the distribution of the hypervolume
measures is not overlapped, meaning a complete outperformance of Case
405 B over Case R in those cases - models a), b), e) and h) -.

• Results disclosed by statistical observations showed that including the

best single-objective solution in the initial population confers a consid-
erable advantage over the case in which the multi-objective optimization
starts off with a random initial population for models a) and b). The low
410 values of the standard deviation of Case B optimums make this issue more
remarkable. However, no evidences were found to sustain this argument
for the remaining models.

• The solution associated with the reference straight barrier has been dom-
inated for some solutions of each optimal front of models exposed to the
415 overall shape optimization, meaning that some designs featuring the same
or even lower overall boundary length and higher shielding efficiency have
been found (see Figures 9- 13).

• The multi-objective optimization performed well to find, generally speak-

ing, wide and uniformly spread-out Pareto-optimal fronts (non-dominated
420 solutions) for the presented barrier models. This is illustrated by the ge-
ometric diversity of noise barriers depicted in Figures 9-16, being most
noticeable, logically, in barrier models featuring overall shape optimiza-
tions - models a) to e) -.

6. Conclusions

425 Results derived of the multi-objective optimization of very thin noise barriers
with perfectly rigid surface were presented. Based on barrier models presented

34
 and studied in a previous work by the authors under single-objective optimiza-
tion, numerical multi-objective analyses were conducted in the pursuit of both
the maximization of the shielding efficiency and the minimization of the cost of
430 all barrier material used, represented by the overall length of the boundaries.
Optimal sets of designs were obtained for each model after running processes
with equal number of fitness function evaluations both when considering a ran-
dom initial population (Case R) and when including the best single-objective
optimal design in the initial population (Case B ). Results from statistical anal-
435 yses showed that evidences revealing the outperformance of Case B over Cases
R, meaning that including the best single-objective solution in the optimiza-
tion run is a real asset, were found in two of the models under study. Fur-
thermore, multi-objective optimizations led to wide and uniformly spread-out
Pareto fronts, which was reflected in the geometric diversity featured by optimal
440 designs.

Acknowledgment

This work was supported by the Ministerio de Economı́a y Competitividad

(MINECO) of Spain and FEDER through research project BIA2014-57640-R.
The authors are grateful for this support.

445 References

[1] Toledo R, Aznárez JJ, Maeso O, Greiner D. Optimization of thin noise bar-
rier designs using Evolutionary Algorithms and a Dual BEM Formulation.
Journal of Sound and Vibration 2015;334:219-238.

[2] Deb K, Pratap A, Agrawal S, Meyarivan T. A fast and elitist multiob-

450 jective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary
Computation 2002;6(2):182-197.

[3] Deb K. Multiobjective Optimization Using Evolutionary Algorithms. Series

in Systems and Optimization; John Wiley & Sons; 2001.

35
 [4] C. A. Coello Coello and D. Van Veldhuizen and G. Lamont . Evolution-
455 ary algorithms for solving multi-objective problems. GENA series; Kluwer
Academic Publishers; 2002.

[5] Srinivas N, Deb K. Multi-objective optimization using non-dominated sort-

ing in genetic algorithms. Tech. Rep.; Department of Mechanical Engineer-
ing, Indian Institute of Technology, Kanput, India; 1993.

460 [6] Srinivas N, Deb K. Multi-objective optimisation using non-dominated sort-

ing in genetic algorithms. Evolutionary Computation 1995;2(3):221-248.

[7] Zitzler E, Thiele L. Multiobjective optimization using Evolutionary Algo-

rithms - a comparative case study. In: Proc. 5th International Conference
on Parallel Problem Solving from Nature (PPSN-V). Amsterdam, Nether-
465 lands; 1998, p. 292-301.

[8] Zitzler E, Thiele L. Multiobjective Evolutionary Algorithms: a compara-

tive case study and the strength pareto approach. IEEE Transactions on
Evolutionary Computation 1999;3(4):257-271.

[9] Coello C. Evolutionary multiobjective optimization: a historical view of

470 the field. IEEE Computational Intelligence Magazine 2006;1(1):28-36.

[10] Coello C, amd D. Van Veldhuizen GL. Evolutionary algorithms for solving
multi-objective problems. 2nd ed.; New York, NY: Springer; 2007.

[11] Grubea S, Jambroić K, Domitrović H. Noise barriers with vary-

ing cross-section optimized by genetic algorithms. Applied Acoustics
475 2012;73(11):1129-1137.

[12] de Lacerda LA, Wrobel LC, Mansur WJ. A dual boundary element for-
mulation for sound porpagation around barriers over an impedance plane.
Journal of Sound and Vibration 1997;202(2):235-247.

[13] Maeso O, Greiner D, Aznárez JJ, Winter G. Design of noise barriers with
480 boundary elements and genetic algorithms. Advances in Boundary Element
Techniques IX, Sevilla, Spain 2008;Keynote Lecture:101-106.

36
 [14] Greiner D, Galván B, Aznárez JJ, Maeso O, Winter G. Robust design
of noise attenuation barriers with Evolutionary multiobjective algorithms
and the Boundary Element Method. In: M. Ehrgott et al. , editor. Lecture
485 Notes in Computer Science; vol. 5467. 2009, p. 261-274.

[15] Greiner D, Aznárez JJ, Maeso O, Winter G. Single- and multi-objective

shape design of Y-noise barriers using Evolutionary computation and
Boundary Elements. Advances in Engineering Software 2010;41(2):368-378.

[16] Deb K, Bandaru S, Greiner D, Gaspar-Cunha A, Tutum CC. An integrated

490 approach to automated innovization for discovering useful design principles:
Case studies from engineering. Applied Soft Computing 2014;15:42-56.

[17] Martin PA, Rizzo FJ, Cruse TA. Smoothness-relaxation strategies for
singular and hypersingular integral equations. International Journal of Nu-
merical Methods in Engineering 1998;42:885-906.

495 [18] Sáez A, Gallego R, Domı́nguez J. Hypersingular quarter-point boundary

elements for crack problems. International Journal for Numerical Methods
in Engineering 1995;38:1681-1701.

[19] Tadeu A, António J, Mendes PA, Godinho L. Sound pressure level atten-
uation provided by thin rigid screens coupled to tall buildings. Journal of
500 Sound and Vibration 2007;304:479-496.

[20] Luo JF, Liu YJ, Berger J. Analysis of two-dimensional thin structures (from
micro- to nano-scales) using the boundary element method. Computational
Mechanics 1998;22(5):404-412.

[21] Bordón JDR, Aznárez JJ, Maeso O. A 2D BEM-FEM approach for time
505 harmonic fluid-structure interaction analysis of thin elastic bodies. Engi-
neering Analysis with Boundary Elements 2014;43:19-29.

[22] Hothersall DC, Chandler-Wilde SN, Hajmirzae MN. Efficiency of single

noise barriers. Journal of Sound and Vibration 1991;146(2):303-322.

37
 [23] Marburg S. Computational Acoustics of Noise propagation in Fluids-Finite
510 and Boundary Elements Methods; chap. Discretization requirements: how
many elements are necessary? Berlin, Heidelberg: Springer; 2008, p. 309-
332. doi:10.1007/978-3-540-77448-8_12.

[24] Telles JCF. A self-adaptative coordinate transformation for efficient nu-

merical evaluation of general boundary element integrals. International
515 Journal for Numerical Methods in Engineering 1987;24:959-973.

[25] Aliabadi MH. The Boundary Element Method. Applications in solids

and structures; vol. 2. West Sussex, UK: John Wiley & Sons; 2002.
doi:10.1002/bate.200301300.

[26] UNE-EN 1793 . Road traffic noise reduction devices. Test method for
520 determining the acoustic performance. Part 3: Normalized traffic noise
spectrum. 1997.

[27] Código Técnico de la Edificación . Documento Básico HR. Protección frente

al ruido. Ministerio de Fomento de España. 2009.

[28] Greiner D, Aznárez JJ, Maeso O, Winter G. Shape design of noise barriers
525 using Evolutionary optimization and Boundary Element. In: In: B. H.
V. Topping, G. Montero, R. Montenegro , editor. Proceedings of the Fifth
International Conference on Engineering Computational Technology. 43;
Stirlingshire, UK: Civil-Comp Press; 2006,.

[29] Lilliefors H. On the Kolmogorov-Smirnov test for nomality with mean

530 and variance unknown. Journal of the American Statistical Association
1967;62:399–402.

[30] Student . The probable error of a mean. Biometrika 1908;6(1):1–25.

[31] Wilcoxon F. Individual comparisons by ranking methods. Biometrics Bul-

letin 1945;1(6):80–3.

38