Embed Size (px)
Citation preview
Expert Systems With Applications 45 (2016) 208–222
Contents lists available at ScienceDirect
Expert Systems With Applications
journal homepage: www.elsevier.com/locate/eswa
Exterior lighting computer-automated design based on multi-criteria
parallel evolutionary algorithm: optimized designs for illumination
quality and energy efficiency
Hugo Rocha a,b,∗, Igor S. Peretta b, Gerson Flávio M. Lima b, Leonardo G. Marques b,Keiji Yamanaka b
a Federal Institute of Goiás, Campus Itumbiara Av. Furnas, 55 - 75524–010 Itumbiara, GO, Brazilb Faculty of Electrical Engineering, Federal University of Uberlândia Av. João Naves, 2121, bl.3N - 38400–902, Uberlândia, MG, Brazil
a r t i c l e i n f o
Keywords:
Computer-automated exterior lighting design
Illumination quality
Energy efficiency
Parallel evolutionary algorithm
Multi-objective optimization
NSGA-II
a b s t r a c t
A proper professional lighting design implies in a continuous search for the best compromise between both
low power consumption and better lighting quality. This search converts this design into a hard to solve
multi-objective optimization problem. Evolutionary algorithms are widely used to attack that type of hard
optimization problems. However, professionals could not benefit from that kind of assistance since evolu-
tionary algorithms have been unexplored by several commercial lighting design computer-aided softwares.
This work proposes a system based on evolutionary algorithms which implement a computer-automated ex-
terior lighting design both adequate to irregular shaped areas and able to respect lighting pole positioning
constraints. The desired lighting design is constructed using a cluster of computers supported by a web client,
turning this application into an efficient and easy tool to reduce project cycles, increase quality of results and
decrease calculation times. This ELCAutoD-EA system consists in a proposal for a parallel multi-objective
evolutionary algorithm to be executed in a cluster of computers with a Java remote client. User must choose
lighting pole heights, allowed lamps and fixtures, as well as the simplified blue print of the area to be illu-
minated, marking the sub-areas with restrictions to pole positioning. The desired average illuminance must
also be informed as well as the accepted tolerance. Based on user informed data, the developed application
uses a dynamic representation of variable size as a chromosome and the cluster executes the evolutionary
algorithm using the Island model paradigm. Achieved solutions comply with the illumination standards re-
quirements and have a strong commitment to lighting quality and power consumption. In the present case
study, the evolved design used 37.5% less power than the reference lighting design provided by a professional
and at the same time ensured a 227.3% better global lighting uniformity. A better lighting quality is achieved
because the proposed system solves multi-objective optimization problems by avoiding power wastes which
are often unclear to a professional lighting engineer in charge of a given project.
© 2015 Elsevier Ltd. All rights reserved.
m
n
t
s
S
g
i
g
1. Introduction
A good lighting design is aimed to enable our environments to
have the best lighting comfort at the lowest possible cost. The lower
is the adaptation efforts of individuals, the greater their sense of com-
fort. From the physiologic perspective, to develop some visual activi-
ties (e.g. reading, identification of objects, localization) our eyes need
specific conditions which depend on the type of activities to be per-
formed (IESNA, 2000). Together with a better illumination quality,
∗ Corresponding author. Tel.: +556434043979.
E-mail addresses: [email protected], [email protected]
(H. Rocha), [email protected] (I.S. Peretta), [email protected] (G.F.M. Lima),
[email protected] (L.G. Marques), [email protected] (K. Yamanaka).
u
T
t
T
http://dx.doi.org/10.1016/j.eswa.2015.09.046
0957-4174/© 2015 Elsevier Ltd. All rights reserved.
ore and more the world demands for energy efficiency and alter-
ative energies, the concept of eco-lighting (Lighting Europe, 2013).
The quality of a given lighting design is currently under the at-
ention of authorities in charge through the adoption of increasingly
tringent standards (European Commission, 2013). As an example,
pain has developed a strategy of savings and efficiency based on a
reat number of actions and regulations guided to improve the Span-
sh power system. One of the main adopted measures was the obli-
ation of improvement of facilities for exterior lighting through eval-
ation and rating of their energy efficiency (Ministerio de Industria,
urismo y Comercio de España, 2008).
Brazilian government developed the ReLuz program which aims
o promote the development of efficient systems for public lighting.
o attend the program, new and more efficient technologies should
H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222 209
b
s
c
d
t
i
i
t
f
a
t
d
l
b
t
l
N
d
s
w
e
(
h
d
l
l
w
p
d
T
n
L
o
p
w
i
G
w
a
E
o
r
a
u
p
C
l
W
t
b
E
e
p
fi
l
d
A
t
a
T
w
s
l
o
r
s
s
p
m
r
r
b
d
2
o
2
i
a
r
t
m
h
p
a
b
m
s
s
T
t
T
o
p
p
t
t
2
m
v
c
l
t
2
r
A
a
i
m
t
e used, regarding principles of energy conservation and technical
tandards in effect, as well as established technical and economical
riteria and procedures (PROCEL, 2004).
To understand the financial impact of an efficient public lighting
esign, consider data from the Company of Electrical Energy Distribu-
ion of Goiás (CELG), one of the important electric energy providers
n Brazil. In 2010, CELG sold 503,617 MWh of energy to supply public
llumination (CELG, 2010) within the state of Goiás (3.99% of Brazilian
erritory). Imagine a scenario where all public illumination facilities
rom Goiás could be exchanged by alternative ones that can achieve
n average of 5% of energy monthly savings. This should mean more
han 3,000,000 USD per year in savings to state coffers.
One of the responsible actors to achieve those goals is the lighting
esign engineer who is defied to specify the best solution to a given
ighting problem. This professional is in a continuous quest to find
etter cost-benefit compromises, considering low energy consump-
ion and better illumination quality. Those requirements convert a
ighting design into a multi-objective optimization problem (Zavala,
ebro, Luna, & Coello, 2014). The aid of specialized softwares is highly
esirable and is considered a requirement to deal with lighting de-
igns. The main problem is that currently available commercial soft-
ares do not include artificial intelligence techniques, leaving to the
ngineer the defy to came up with a feasible solution.
Extending the well known Computer-Aided Design
CAD) (Horváth & Vroom, 2015), Kamentsky and Liu (1963) probably
ave coined the term Computer-Automated Design (CAutoD) when
escribing a computer program built to design character recognition
ogic based on the processing of data samples which searches for
ogic circuits having certain constraints on hardware design. This
orks therefore adopts the term CAutoD to designate any design
erformed by computers, without human interference besides prior
efinitions of available materials and problem-related constraints.
hose automated designs have been successfully achieved by tech-
iques such as evolutionary computation (e.g. Hornby, Lohn, and
inden (2011)), among others.
Multi-objective optimization problems are considered a hard class
f problems to solve. Evolutionary algorithms (EA) are a class of com-
utational heuristics that try to mimic Darwinian evolution and are
idely used to successfully solve those class of problems. Interest-
ng to mention is, as stated by Rabaza, Pena-Garcia, Perez-Ocon, and
omez-Lorente (2013), “nevertheless, such algorithms have not been
idely applied to lighting.”
This work proposes a system based on evolutionary algorithm
imed to computer automated exterior lighting design, named as
LCAutoD-EA (Exterior Lighting Computer-Automated Design based
n Evolutionary Algorithm), suitable for irregular shaped areas with
estrictions on placement of lighting poles. The proposed system is
variant of genetic algorithms (Goldberg, 1989; Holland, 1975) and
ses a variable-size genotype to represent a candidate solution.
After a proper literature survey, only three papers that could ap-
roach the subject of this work have been found, so as to be known:
astro, del Acebo, and Sbert (2012) worked with EA to design indoors
ighting with soft geometrical constraints (obstacles); Brownlee and
right (2015) presented EA aimed to design building floors in order
o reduce energy demand, where lighting is just one of the aspects to
e considered; Carlucci, Cattarin, Causone, and Pagliano (2015) used
A to design building variants to match the European near zero en-
rgy buildings directive, where again lighting is just one of the as-
ects to be considered. Within this scenario, this present work is the
rst to use EA techniques to perform computer-automated exterior
ighting design respecting both hard geometrical constraints (forbid-
en positioning in irregular shaped areas) and applicable standards.
uthors have no knowledge about other works intended as so.
Due to the high dimensionality of a lighting design problem,
he utilization of high-performance computation is needed. This is
chieved by using a cluster of computers and parallel programming.
he evolved design for a given lighting problem is then achieved
ithin a relatively small period of time.
Section 2 shows some concepts on which this work is based, as:
tandards taken into account; some metrics to compare different
ighting designs; the point-by-point method to calculate illuminance
n a meshed working plane; the almost automated lumen method for
egular areas; and some basics on genetic algorithms with some in-
ights from the parallel variant and from the adopted multi-objective
trategy. Section 3 presents the ELCAutoD-EA system, detailing pro-
osals on the chromosome representation, genetic operators, fitness
easurement and used computational environment. Section 4 shows
esults and discussions when using the proposed system for a simple
egular area example and for an irregular area from a real parking lot,
oth compared to other methods. Section 5 presents conclusions and
irections for ongoing work.
. Theory
This section presents some concepts that guided the development
f the proposed system.
.1. Standards
The main objective for a standard addressed to lighting designs
s to set out the necessary lighting requirements in order to ensure
n acceptable minimum level of safety and comfort to users of the
especting artificial lighting environments. This work proposes that
he evolved projects by the ELCAutoD-EA system must accord to the
ain international standards in this area.
The Illuminating Engineering Society of North America (IESNA)
as several publications on this theme being that way the most im-
ortant standard reference recognized in illumination science and
pplication (IESNA, 2000). Concepts addressed in the Lighting Hand-
ook Reference & Application were widely referred to in the develop-
ent of this work.
Another important reference in this field is the European
tandards from which we address to the Spanish current ver-
ion (Ministerio de Industria, Turismo y Comercio de España, 2008).
hose standards have the goal of establish technical conditions for
he design, execution and maintenance of exterior lighting facilities.
hey are also concerned about, among other issues, the improvement
f energy efficiency and savings as well as the limitation of the light
ollution and reduction of intrusive or annoying light.
The Brazilian ABNT NBR 5101 (ABNT, 2012), which rules about
ublic illumination facilities, fix some requirements considered to be
he minimum necessary to illuminate public ways destined to the
raffic of pedestrians and vehicles.
.2. Metrics from standards
A lighting design to be considered must be based on the compro-
ise between the illumination quality and the energy efficiency pro-
ided by the desired facility. For this purpose, this work detach two
oncepts that can be used to compare performances from different
ighting designs: the illuminance uniformity (global and general) and
he energy efficiency rate.
.2.1. Illuminance uniformity
Illuminance uniformity is a calculated quantity that intends to
epresent the quality of the illuminance distribution on a given area.
good level of uniformity is desired to avoid strong shadow regions
nd to ensure the comfort and safety to practices that will be hosted
n the area to be illuminated.
This uniformity could be achieved by some rates of minimum,
aximum and average illuminance values calculated on the respec-
ive meshed working plane from the area of interest. Literature shows
210 H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222
U
U
Table 1
Energy efficiency label for a lighting facility (repro-
duced from (Ministerio de Industria, Turismo y Com-
ercio de España, 2008)).
Energy class Energy consumption index
A ECI < 0.91
B 0.91 ≤ ECI < 1.09
C 1.09 ≤ ECI < 1.35
D 1.35 ≤ ECI < 1.79
E 1.79 ≤ ECI < 2.63
F 2.63 ≤ ECI < 5.00
G ECI ≥ 5.00
Pn
γn
l1 l2
ln
dn
Fig. 1. Guideline to geometric quantities.
2
i
f
a
m
o
c
a
fi
l
t
l
m
b
t
o
F
n
d
p
s
s
E
w
n
b
E
w
t
o
E
t
i
p
some different ways to calculate this measurement. The most widely
used is known as global uniformity U0 (ABNT, 2012; IESNA, 2000;
Rabaza et al., 2013), which is the rate of the minimum by the aver-
age of illuminance values and is presented in Eq. (1). Another one is
the general (or extreme) uniformity Ug, which is the rate of the min-
imum by the maximum of illuminances values (Ministerio de Indus-
tria, Turismo y Comercio de España, 2008) and is presented in Eq. (2).
Note that (IESNA, 2000) also mention the possibility of calculating
the inverse of the general uniformity (rate of the maximum by the
minimum).
0 = Emin
Eav(1)
g = Emin
Emax(2)
where Emin is the minimum value calculated for the illuminance on
the meshed working plane; Eav is the average of illuminance values;
and Emax is the maximum of illuminance values.
To exterior public spaces, standards recommend an average of 30
lx. Greater levels of illuminance could be desirable when the area in
question is supposed to host visual tasks which require a greater per-
ception of details. Also, the variation of illuminance levels should not
be greater than 1/3 of the ideal average, more or less (IESNA, 2000).
For example, the International Commission on Illumination
(CIE) (European Commission, 2013) states that an exterior lighting
design should have a minimum average illuminance of 20 lx and a
minimum global uniformity of 0.4 (besides that for some few area
purposes, it could be 0.35). Brazilian ABNT also state those as minima
values.
2.2.2. Energy efficiency rating
The energy efficiency (ε) of a lighting facility could be calculated
by Eq. (3):
ε = A × Eav
Pow(3)
where A is the surface area to be illuminated [m2]; Eav is the average
illuminance [lx]; and Pow is the total of installed electrical power [W].
The energy efficiency rating (Iε) of a lighting facility is then de-
fined as the ratio of its energy efficiency by the energy efficiency ref-
erence value (εR) with respect to the average illuminance achieved
by the lighting design, as shown in Eq. (4). This reference value
can be obtained from (Ministerio de Industria, Turismo y Comercio
de España, 2008, ea-01, table 3) .
Iε = ε
εR
(4)
To simplify the data interpretation of the energy efficiency
rating, a classification system was developed. The Spanish stan-
dards (Ministerio de Industria, Turismo y Comercio de España, 2008)
presents an energy consumption labeling scheme aligned with the
classification labels first designed for household appliances. It is de-
fined as a scale of seven levels, starting with class A (a more efficient
facility with a lower energy consumption) until class G (a less effi-
cient facility with a greater energy consumption). The energy con-
sumption index (ECI) used to locate which class a facility belongs
to is equal to the inverse of the energy efficiency rating, as seen in
Eq. (5).
ECI = 1
Iε(5)
After calculating the ECI for a given lighting design, Table 1 gives
its energy class.
.3. Point-by-point method
The point-by-point calculation method (a type of directed light-
ng method) is the most often used method to calculate illumination
or external lighting designs due to the supposed absence of greater
rtifacts (like walls or ceilings) responsible for reflecting light. This
ethod is also recommended by the Illuminating Engineering Society
f North America (IESNA). When applying this method, the designer
ould obtain the luminance provided by one or more light sources at
ny desired point from a meshed working plane. A light source is de-
ned by a given luminaire which could be understand as a complete
ighting unit consisting of one or more lamps which are connected
o a power source and a reflector aimed to direct and distribute the
ight. This work uses this method to calculate illuminance.
Part of the luminous flux emitted by lamps is absorbed by the lu-
inaire and does not contribute to the ambient lighting. The flow
alance is spent above and below a horizontal plane passing through
he center of the luminaire. The beams of light irradiated directly
n the work plane mainly contributes to the achieved luminance.
or each luminaire, the photometric data used for calculating lumi-
ance are obtained by their respective IES files (an international stan-
ard protocol from the Illuminating Engineering Society to describe
hotometric data computer files) which are used in many existing
oftwares.
The total illuminance, given by adding contributions of each light
ource, is represented by Eq. (6):
total =N∑
i=1
Ei (6)
here Ei is the illuminance component of the source i; and N is the
umber of light sources.
The illuminance Ei, held by each light source individually, is given
y Eq. (7) which has related quantities presented in Fig. 1:
i = IPi· cosγ
d2(7)
here IPiis the intensity of the light source i toward the point P; γ is
he angle between the normal to the working plane and the direction
f the considered point P with respect to the light source; and d is the
uclidean distance between the light source and the point P.
Regarding Fig. 1, �N is the normal to the plane; γ i is the angle be-
ween �N and the segment LPi; h is the luminaire height of fixation; lis the horizontal distance between the light source (represented by
oint L) and the point P of interest; d is the Euclidean distance from
i i
H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222 211
t
o
w
i
f
d
a
m
l
t
i
c
a
2
fl
fl
a
d
r
(
c
b
n
A
g
l
s
p
E
K
w
r
m
a
N
w
a
i
fl
b
F
a
l
t
v
t
e
2
r
a
m
u
a
u
l
s
e
p
r
n
2
e
l
l
i
t
l
t
m
l
c
r
2
e
l
t
i
s
j
t
i
P
P
M
d
e
d
fl
2
o
v
d
t
s
i
i
d
2
d
u
t
he light source to the point Pi; i = 1, 2, · · · n; and n is the number
f interest points on the meshed working plane (discretized space)
hich constitutes the area to be illuminated.
The flux intensity emitted by a light source in a certain direction
s given by the photometric distribution of the light source and varies
or each type of luminaire (and respective lamp) used. The use of
igitized curves through photometric data standard files (IES) allows
utomatic light intensities calculations, simplifying computer imple-
entations.
The point-by-point method needs the initial configuration of
ighting installation (identified light sources with respective posi-
ions in the horizontal plane, pole heights and luminaires) to provide
lluminance values for each point of the meshed area of interest. This
ould be performed just on the working plane (2D) or on several par-
llel planes between the light sources and the working plane (3D).
.4. Lumen method
The lumen method (IESNA, 2000) considers the total of luminous
ux needed to illuminate a ambient environment. Room surface re-
ectances (ceiling, walls and floor) are considered in this method
s well as the desired equipment (lamps, fixtures and ballasts). To
istribute uniformly the illumination on a given rectangular envi-
onment with diffuse reflection surfaces, just one type of luminaire
desired equipment) is taken into account. This method enables a
ertain level of automated design because it can calculate the num-
er of light sources necessary to achieve the desired average illumi-
ance when placing those luminaires equally apart from each other.
ctually, some commercial softwares make use of this method to
et some level of automation (Philips LiDAC, 2011). This method is
imited to rectangular shaped areas, though, wherein needed light
ources could be placed at will. This work uses this method for com-
arisons only.
Starting with the dimensions of the area, the lumen method uses
q. (8) to achieve the room index K.
= c · l
h · (c + l)(8)
here c, l and h are the width, length and height of the room [m],
espectively. h is also known as the vertical distance between the lu-
inaires and the working plane.
The number of luminaires to a given ambient environment is then
chieved by Eq. (9).
l = Eav · A
nl · φn · U · MF · BF(9)
here Nl is the required number of luminaires; Eav is the specified
verage illuminance [lx]; A is the surface area of the room [m2]; nl
s the number of lamps in each luminaire; φn is the nominal lamp
ux; U is the utilization factor; MF is the maintenance factor; BF is the
allast factor. To specify U, values presented in (OSRAM, 2013, pp. 22,
ig. 27) must be interpolated. MF should be considered between 0.6
nd 0.8, according to (OSRAM, 2013, Section 4.5.4), and BF is generally
ess than one, unless it is the ballast of reference (in this case, BF = 1).
The lumen method is widely used for indoor environments. Note
hat it can also be used when dealing with exterior (or outdoor) en-
ironments, once Table (OSRAM, 2013, pp. 22, Fig. 27) allows calcula-
ions by providing the utilisation factor for a room surface reflectance
qual to zero.
.5. Evolutionary algorithm
The kernel of the proposed system is an evolutionary algo-
ithm based on Genetic Algorithms (GA) from (Holland, 1975)
nd (Goldberg, 1989). A genetic algorithm generates solution to opti-
ization problems using techniques inspired by the Darwinian nat-
ral evolution, as mutations, selection and crossovers occurring in
population of individuals. Within a GA population, each individ-
al represents a candidate for the solution of the investigated prob-
em (Goldberg, 1989). As in GA framework, the initial population
tarts with a randomly selected individuals and, through genetic op-
rators, each new generation of individuals evolve towards higher ex-
ected fitness (or better quality solutions). The algorithm ends when
eaching a certain termination criterium (for example, a maximum
umber of generations).
.5.1. Parallel evolutionary algorithm
The parallel variant of an evolutionary algorithm could have sev-
ral architectures. The Coarse-Grained model (also known as the Is-
and model) divides the entire population into multiple local popu-
ations (sub-populations) and each instance of the algorithm (each
sland) is then executed in parallel. This model also integrate rou-
ines to enable migration of individuals between the available is-
ands (Luque & Alba, 2011; Shenfield & Fleming, 2014).
The topology for communication between different instances of
he GA (sub-populations) is part of the migration strategy and is
apped upon some physical network. Migration frequency, rates, se-
ection strategy and topology are parameters that has to be carefully
hosen, due to the high influence in the parallel evolutionary algo-
ithm performance (Kimovski, Ortega, Ortiz, & Baños, 2015).
.6. NSGA-II
Multi-objective evolutionary algorithms (Zavala et al., 2014; Zhou
t al., 2011) are heuristics widely used to deal with optimization prob-
ems whose set of feasible solutions consists of elements which fulfill
wo or more objectives. When dealing with multi-objective problems,
t is a common knowledge that sometimes is impossible to optimize
everal objectives simultaneously, as often the fulfill of a single ob-
ective could lead to the derangement of the others. For example, in
his work the search for a good energy efficiency can easily impair the
llumination quality of an area.
The widely recognized multi-objective strategy is NSGA-II (Deb,
ratap, Agarwal, & Meyarivan, 2002; Kamjoo, Maheri, Dizqah, &
utrus, 2015; Rout & Acharya, 2015; Shaygan, Alimohammadi,
ansourian, Govara, & Kalami, 2014). Using the concept of Pareto’s
ominance and Pareto-optimal fronts, this algorithm provides an
fficient scheme to sort individual fitness in order to enable the
ecision of when an individual is better than another. The NSGA-II
owchart is presented in Fig. 2.
The NSGA-II sorting algorithm has two main steps (Deb et al.,
002): the fast non-dominated sort which distributes the population
f candidate solutions into separated fronts with respect to each indi-
idual fitness for each objective (see Section 2.6.1); and the crowding-
istance which avoids the premature convergence of the population
hrough a scheme to maintain its diversity. This is achieved by as-
igning to each individual from the same front a value with respect to
ts respective degree of isolation. The more isolated is the individual
n a given front, the better it is classified. An example of crowding-
istance calculation is given in Fig. 3.
The overall NSGA-II procedure is presented in Fig. 4.
.6.1. Non-dominated sorting algorithm
In the fast non-dominated sort scheme, to each individual (candi-
ate solution) p, within a population of candidates P of size T individ-
als, are determined the following:
• np: the number of candidates in P which dominates p (Pareto’s
dominance),• Sp: the set of candidates in P which are dominated by p, and• f: the front to which p belongs.
First, the set P is traversed and each individual p ∈ P is compared
o the others aiming to determine the quantity of individuals that
212 H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222
Fig. 2. Flowchart for NSGA-II.
Fig. 3. Crowding-distance calculation. Points marked in filled circles are solutions of
the same non-dominated front (adapted from (Deb et al., 2002)).
Fig. 4. NSGA-II procedure (adapted from (Deb et al., 2002)).
o
(
w
f
e
f
dominate p (np) and to identify individuals dominated by p (regis-
tered in Sp). In the end, each individual p with np = 0 is placed within
front#1.
Each individual p which is placed within front#f (the Pareto-
ptimal ones are in f = 1) has a set of dominated individuals
S − p). All individual q ∈ Sp has nq decremented. In the case
here nq = 0, the respective individual q is then placed within
ront# f + 1. This procedure is performed recursively, until ev-
ry individual in the original population is assigned to some
ront.
H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222 213
Table 2
Lamps database for exterior lighting design (lamp wattage and
luminous flux are rated at EM 25 °C). All listed lamps are from
OSRAM.
Product name Lamp wattage Luminous flux
NAV-E 100W SUPER 4Y 100 10200
NAV-E 150W SUPER 4Y 150 17000
NAV-E 250W SUPER 4Y 250 31100
NAV-E 400W SUPER 4Y 400 55500
3
n
a
i
r
t
d
c
3
c
c
r
b
l
b
t
d
c
i
t
r
a
m
s
a
r
i
u
c
t
T
d
m
3
c
S
t
f
l
a
i
a
t
h
o
t
z
f
t
d
a
t
r
a
I
w
a
φa
t
l
p
3
l
a
a
c
o
e
s
e
t
s
v
t
t
b
E
b
r
h
o
t
i
3
. Method
The core of the proposed system (the evolutionary algorithm ker-
el) is a variant of genetic algorithms (presented in (Holland, 1975)
nd goldbergGA1989) by changing the same-size vector idea of an
ndividual by a vector of lists which is the data structure that rep-
esents the genotype of a candidate for the desired design. This sys-
em also uses NSGA-II as a sorting algorithm to evolve multi-objective
esigns and the island model (Coarse-Grained) to perform parallel
omputations.
.1. Inputs to exterior lighting CAutoD
Exterior lighting design could be simplified by the appropriate
hoice of which and how many luminaires will be installed (in-
luding fixture and lamp), which is going to be the height of each
espective lighting pole, and where each of those luminaires will
e placed. Regarding unusual areas, some placement restrictions on
ighting poles must be respected. Lighting standards in force must
e respected. The financial dimension must be kept in considera-
ion, i.e., keeping the maintenance cost as low as possible is highly
esirable.
In order to simplify the data input to the proposed system, a Java
lient was developed. User needs to previously register some basic
nformation as:
• available lighting poles and luminaires (fixtures and lamps);• a simplified layout of the (exterior) area to be illuminated which
includes subareas where the placement of lighting poles is al-
lowed, as can be seen in Fig. 5b;• respective isolux curves for each of the available luminaires1.
The Java client (screenshots in Fig. 5) loads the area layout as
wo main types of polygons defined by the user: the ones which
epresent where it is allowed to place lighting poles (middle gray
rea, see Fig. 5b), and the ones which represent where the place-
ent is forbidden (light gray areas). The outside area which is not
upposed to be lightened is also taken into account (dark gray
rea) – any considerable illumination that reaches it is a loss of
esources.
The configuration screen (seen in Fig. 5c) contains previously reg-
stered luminaires, pole heights, target exterior area, and desired val-
es for illuminance. Note that currently the luminaire is a hypotheti-
al one (see Section 3.1.1). All of them can be chosen depending on
heir availability and intention to the desired design. Presented in
able 2 are the default lamps already available in the Java client’s
atabase. The user can choose from database or register any other
aterial he desires.
.1.1. Hypothetical luminaire
Manufacturers of luminaires usually provide the photometric
urves related to their catalog products. Some of them also provide
1 Currently only an isotropic isolux curve is available for all luminaires (see
ection 3.1.1).
o
c
hrough their website the related photometric data files in the IES
ormat. Using the photometric data from a specific luminaire, the il-
uminance of any point on the working plane can be calculated. Actu-
lly, those files provide values of luminous intensity in discrete angles
ncrement, but it is possible to calculate a value for an intermediate
ngle using strategies for interpolation.
To avoid interpolations and as a part of the simplification efforts,
he only type of fixture considered at this stage of the research is a
ypothetical one, designed as a direct lighting luminaire (100% of its
utput is directed downwards). The luminous intensity distribution
akes the form of a sphere, with the luminaire located on the sphere
enith. This distribution therefore can be described by a continuous
unction. Analyzing planes parallel to the working plane, can be found
hat this hypothetical luminaire has an isotropic luminous intensity
istribution. In planes perpendicular to the working plane, it has the
ppearance of a circle. The IES-like photometric curve for this hypo-
hetical luminaire is shown in Fig. 6.
When mounting a given lamp to this hypothetical luminaire, its
espective luminous intensity I, in candelas, are given in function of
ngle γ :
(γ ,φ) = �
4π
(1 + cos 2γ
cos γ
)(10)
here �, in lumens, is the luminous flux; γ is the elevation or vertical
ngle; and φ is the horizontal angle. Note that I is independent of
, i.e., it is isotropic through the planes parallel to the illuminated
rea. The maximum luminous intensity is given in the direction of
he luminaire focus, i.e., where γ = 0 rad. Because this hypothetical
uminaire can be only installed parallel to the ground, on the working
lane it is assumed to be the point straightly below it.
.2. Chromosome
The solution candidate, also known as the individual of a popu-
ation in the framework of evolutionary algorithm, is represented by
set of unknown values (its genotype). The desired system must be
ble to translate from this representation to a proper design specifi-
ation (its phenotype).
In lighting designs, there is no previously known proper number
f light sources, besides the fact that lighting engineers are able to
stimate this number. An adequate representation is essential to the
uccess (or the failure, if it is missing) of any evolutionary algorithm,
specially when aiming for computer automated design. The problem
o be solved here, exterior lighting, needs the flexibility of a repre-
entation that has not a fixed size. This work uses a chromosome of
ariable size to allow a better performance in an undetermined mul-
idimensional search space.
The proposed chromosome is therefore a vector of lists. The vec-
or has a fixed size and each index of its elements represents a possi-
le combination from enlisted pole heights and available luminaires.
ach element is associated with a list which contains a variable num-
er of elements (that could be none). Each of those list elements
epresents the coordinates (x, y) for positioning a given pair of pole
eight and luminaire. Those discretized coordinates belongs to a set
f possible coordinates built from the “allowed position” polygons
aken from the area layout. An example of this chromosome is shown
n Fig. 7.
.3. Genetic operators
Regarding the proposed chromosome representation, new genetic
perators must be defined. This work also presents the proposed
rossover and mutation operators.
214 H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222
Fig. 5. Some screenshots from the Java client.
t
a
a
3
l
s
s
3.3.1. Crossover
Crossover (also known as recombination) is the genetic operator
capable of converging a population of individuals towards the most
fit ones. It is responsible for exploiting the search space. After select-
ing two individuals from the population (in this case by the strategy
of tournament), the crossover algorithm runs through each vector el-
ement and chooses randomly one crossover point from their respec-
tive list of coordinates. The generated offspring is going to combine
the first list elements from one parent with the last list elements from
he other parent, with respect to the order in the main vector. An ex-
mple of crossover is presented, where the parents are shown in Fig. 8
nd the offspring is shown in Fig. 9.
.3.2. Mutation
Mutation is the genetic operator capable of diverging a popu-
ation of individuals, as divergence of individuals is extremely de-
irable inside the framework of evolutionary algorithms. It is re-
ponsible for exploring the search space, looking for other optimum
H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222 215
50100
150200
30º
60º
90º-90º
-60º
-30º
0º
0º
30º
60º
90º
Fig. 6. Photometric curve of the hypothetical luminaire, units in cd/1000 lm.
Fig. 7. Chromosome c. The vector has a fixed size equal to nm and index k. Poln + Lumm
represents the pair of pole height n and luminaire m. (x(k)nm, y(k)
nm) is the coordinate of the
kth pair 〈n, m〉 located on the area layout. .
Table 3
Possibilities for mutation: (√
) affected entity; (✗) non-affected entity. Note that once is
determined to occur a mutation, the proper variant is chosen randomly with uniform
distribution.
Mutation operator variants
Variant # 1 2 3 4 5 6 7
Pole height√
✗√
✗√
✗√
Lighting pole position ✗√ √
✗ ✗√ √
Luminaire ✗ ✗ ✗√ √ √ √
s
l
t
l
d
c
e
3
s
S
i
t
t
s
c
a
b
F
olutions. There are some different proposed mutation operators, al-
owing diverse mutations to occur. This operator can exchange any of
he following entities values to any other one available: pole height,
uminaire, and lighting pole position. Those entities are changed ran-
omly inside each selected individual, according to Table 3. The per-
entage of individuals subject to be mutated is defined within the
volutionary algorithm parameters before its execution.
Fig. 8. Crossover: two chosen parents (∗’s indi
.4. Fitness evaluation
The proposed system uses NSGA-II, a multi-objective sorting
cheme, to address the relative fitness of each of the individuals (see
ection 2.6). The implemented NSGA-II deals with two objectives: an
llumination criterion and an energy efficiency criterion. The first cri-
erion reflects the aim to distribute uniformly the illumination over
he exterior area inside some interval of illuminance determined by
tandards. This is implemented as follows:
• First, the fitness function must calculate the illuminance of each
point on the meshed working plane. To achieve this, the point-by-
point method (see Section 2.3) is used for a given individual which
provides information on light sources.• Then, a conditioning expression is applied to each calculated illu-
minance value. This expression has the form presented in Eq. (11)
and it is based on the Gaussian distribution of Statistics (see
Fig. 10). The idea is to use the illuminance required by the stan-
dards to be the mean value, using the accepted tolerance as stan-
dard deviation value.
fit(i) = 1√2πσ 2
exp
[− (Ei − μ)
2
2σ 2
](11)
where fit(i) is the conditioned value calculated from the illumi-
nance value of a single point i on the mesh; Ei is the illuminance
value on point i as in Eq. (7); μ is the desired standard illuminance
value (for example, 30 lx); and σ is the standard deviation value
to this “Gaussian-like” distribution (for example, 10 lx). Note that,
in this example, illuminance values between 20 and 40 lx will be
better evaluated over values outside of this interval, and that the
best evaluated value for illuminace is the desired value of 30 lx.• After applying the conditioning expression over illuminances, the
individual j has fitness with respect to this criterion F1 equal to the
summation of all conditioned illuminances, as seen in Eq. (12).
F 1( j) =N−1∑i=0
fit(i) (12)
where F1(j) is the fitness of the individual j for the criterion of
uniform illumination; N is the total number of points from the
meshed working plane; and fit(i) is the conditioned illuminance
of each point i from the mesh.
Note that the implemented NSGA-II is configured to maximize this
riterion of uniform illumination.
The second criterion is about energy efficiency. This is done by
dding up each electrical power consumption from all lamps encoded
y a given individual, as seen in Eq. (13).
2( j) =NL−1∑i=0
e j(i) (13)
cate randomly chosen crossover points).
216 H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222
Fig. 9. Crossover between parents p1 and p2 from Fig. 8 generates offspring o1 and o2.
0 10 20 30 40 50 600
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
Region of OptimalIlluminance
Illuminance [lx]
f(i)
Fig. 10. Conditioning expression with respect to the calculated illuminace of a given point on the working plane mesh.
a
a
p
A
w
fi
F
c
i
fi
e
a
e
t
o
3
d
where F2(j) is the fitness of the individual j for the criterion of energy
efficiency; NL is the total number of lamps encoded by a given indi-
vidual; and ej(i) is the electrical power consumption from the lamp
i enconded by individual j. The implemented NSGA-II is configured
to minimize this criterion of energy efficiency. Note that both criteria
are concurrent, i.e., the overall optimization problem needs to deal
with minimization and maximization to solve the problem of a good
uniform distribution of illumination with the minimum of electrical
power consumption.
3.5. Choosing a final design
An evolutionary algorithm using NSGA-II does not reaches a best
individual, but a set of better individuals with different balance be-
tween performance in each one of both criteria. This set is referred
as front#1 (see Section 2.6) because of Pareto optimal solutions the-
ory. Besides the fact that, for a given problem, an infinite numbers of
solutions could exist that fits Pareto optimal (or nondominated) defi-
nition, a system that is supposed to do exterior lighting CAutoD needs
to infer a specific design to be the final one.
To achieve this “elected” design, a metric was build to integrate
performance in all criteria. Note that this overall performance metric
is applied only to individuals (design solutions) present in front#1
t the end of the evolutionary algorithm run. The proposed metric,
s presented in Eq. (14), is pretty straightforward when dealing with
roblems with both minimization and maximization criteria.
v(i) = F 1(i)
max(F 1(1 : N))+ min(F 2(1 : N))
F 2(i)(14)
here Av(i) is the proposed metric applied to individual i; F1(i) is the
tness of the individual i for the criterion of uniform illumination;2(i) is the fitness of the individual i for the criterion of energy effi-
iency; max(F1(1: N)) is the maximum fitness F1 found between the N
ndividuals which belongs to front#1; min(F2(1: N)) is the minimum
tness F2 found between the N individuals which belongs to front#1.
Shown in Fig. 11 is the front#1 individual fitness (F1 × F2) in the
nd of a system run example. Fig. 12 presents sorted values obtained
fter applying this metric to all individuals in front#1 in the same
xample. The proposed system will elect as the winner design one of
he individuals from the front#1 which has the greater value on this
verall performance metric.
.6. Evolutionary algorithm parameters
To execute the system, the following parameters are suggested by
efault, even that experienced user can change some of them at will:
H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222 217
Fig. 11. Fronteira 1.
Fig. 12. Sorted values obtained from application of the overall performance metric.
a
u
3
l
t
t
s
l
• Parallel model : Coarse-Grained (also known as island model);• Migration model : random (all islands connected) – migration rate
of 1% occurring at each 10 generations;• Crossover probability : 85%;• Mutation probability : 15%;• Maximum number of generations : 50;• Size of population : 20,000 (divided into 500 individuals per is-
land);• Fitness sorting strategy : Non-dominating crowding distance;• Parent selection strategy : tournament with 1% of individuals (5
by default);• Reduction step strategy : survival of best individuals.
Note that when a mutation operator takes place (with 15% prob-
bility by default), it will perform changes on individuals using an
niformly random strategy variant picked from Table 3.
.7. Computational environment
The proposed system is constituted of a Java client GUI (user
ayer) and an evolutionary algorithm implemented in C++ (applica-
ion layer) running in a cluster of computers. The communication be-
ween those layers is done through HTTP protocol (a dedicated web
ervice/client) which enables remote access to the system. The user
ayer can run in any machine connected to the Internet that can run
218 H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222
Fig. 13. N MPI processes working in M nodes.
Fig. 14. A square area of 1,600 m2 to be illuminated.
Table 4
Comparison between the lumen method and the ELCAutoD-EA (best values of each
column in bold).
Designed by Electrical power
consumption
(kWh)
Average
illuminance
Uniformity Energy class
U0 Ug
Lumen method 0.9 26.85 lx 0.40 0.29 A (0.63)
ELCAutoD-EA 1.1 26.40 lx 0.47 0.35 A (0.78)
a
r
l
L
a
e
fl
t
K
w
h
t
f
i
f
n
s
N
w
d
e
o
d
p
i
s
n
Java programs. It is presented in Section 3.1. The application layer
runs in a computer cluster (Beowulf type) formed by five intercon-
nected computers Xeon E5504 (Nehalem-EP) with 8 cores @2.0 GHz
each (40 cores total). The installed OS is the Linux Ubuntu 10.04LTS.
The application uses Message Passing Interface (MPI), more specif-
ically the C-library MPICH2 (MPICH), to do the parallel processing
through message exchanging (Kirby & Karniadakis, 2003). This idea
is illustrated in Fig. 13.
The option for a cluster of computers instead of other paralleliza-
tion schemes (e.g. multithreading, GPGPU/CUDA, GPGPU/OpenCL)
was due to the best possible compromise between high performance
computational power, available budget and reliability at the time this
research has begun (year of 2010). The cluster of computers were al-
ready available at that moment.
An island-model parallelization scheme is known to present up
to supra-linear speed ups (Collet, Krüger, & Maitre, 2013). To imple-
ment the Coarse-Grained (island) model for the proposed parallel
evolutionary algorithm, the application generates 48 processes with
one instance of evolutionary algorithm each, i.e., each process repre-
sents an “island” with independent populations. Once in a while the
islands exchange individuals through migrations. The main process
is the process #0, designated to control the parallel execution. Also,
process #0 always runs in the front-end machine of the cluster which
means that it can directly access all input/output hardware, as well
as communicate to the web service (indirectly, with the Java client)
through text files.
4. Results and discussion
This section presents some example results. First, the lighting de-
sign for a simple hypothetical area given by the lumen method (see
Section 2.4) in comparison to the design evolved by the proposed sys-
tem; and a real lighting design made by a specialized engineer in-
stalled on a parking lot in comparison to the design evolved by the
proposed system.
4.1. A simple example
A simple example is present here, an square exterior area of
1,600 m2 (40 × 40 m) to be illuminated presented in Fig. 14. The
aim is to illuminate mainly inside the area of interest (minimizing the
outside area which is shown in dark gray) respecting standards con-
strains and respecting the allowance (gray subareas) or forfeit (light
gray subareas) of lighting pole positioning.
To do the automated exterior lighting design, two methods were
used for the sake of comparison: the lumen method and the pro-
posed system. Note that the lumen method is a method to equidis-
tantly place light sources through a given regular area. The result of
this method could be seen as the ideal positioning for a set of lighting
poles within regular areas, assuming that all area is available to in-
stall light sources. It is fairly to see that the ELCAutoD-EA system has
n obvious advantage over the lumen method: the former will always
espect geometrical constrains for lighting pole placements while the
atter will not.
umen method. Considering its limitations, the lumen method was
pplied to the area of interest. First, external areas could be consid-
red as not having light reflexions due to ceilings, walls or floor (re-
exion rates equal to zero) (IESNA, 2000). Then, in reference to Eq. (8),
he room index K should be:
= 40 · 40
10 · (40 + 40)= 2
here the area has length and width both equal to 40 m; and the pole
eight is assumed to be 10 m.
According to (OSRAM, 2013, pp.22, Fig. 27), for a K = 2, the utiliza-
ion factor U = 0.5. Note that both maintenance factor MF and ballast
actor BF are ignored (considered equal to 1) which means a recently
nstalled facility using ballasts of reference. The option to ignore those
actors does not cause losses to the comparison, once they are also ig-
ored when using the proposed system.
Finally, in reference to Eq. (9), the indicated number of light
ources needed to illuminate this area should be:
l = 30 · 1, 600
10, 200 · 0.5= 9.4
here the desired average illuminance is 30 lx (as stated by stan-
ards); and lamps are considered to provide 10,200 lm each.
The lumen method applied for this hypothetical area suggests an
quidistant distribution of 9.4 light sources. Considering that the area
f interest is a square, the only possibility in this case is to round
own this number to 9. This means to place 3 lines of 3 lighting
oles each, equidistants one from another. The resulting isolux curve
s shown in Fig. 15. Some comparison metrics for this design are pre-
ented as part of Table 4.
As part of the lack of geometric constraints, the lumen method is
ot supposed to be used for irregular shaped (non-rectangular) areas.
H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222 219
Fig. 15. Isolux curve from design guided by the lumen method. Note that there are
some lighting poles placed on forbidden areas.
Fig. 16. Isolux curve from design evolved by the ELCAutoD-EA.
A
a
E
s
r
i
t
s
(
u
t
S
t
p
l
s
a
E
o
c
c
p
t
b
a
t
s
i
t
o
4
w
t
e
i
p
o
a
i
a
l
g
d
l
t
h
t
t
o
s
a
t
b
f
s
F
i
E
w
s
a
d
m
r
v
d
g
3
(
g
i
p
t
lso, note in Fig. 15 that lighting poles identified by letters A, C, E, G,
nd I needed to be placed on forbidden subareas.
LCAutoD-EA. The user of the proposed system must have some ba-
ic data-base which includes pole heights, luminaires (with their
espective photometric curves), and a simple layout of the area of
nterest (including subareas allowed to receive lighting poles and
he ones where this is forbidden). Note that at this point of the re-
earch, the system works with an isotropic photometric curve only
see Section 3.1.1).
Note that the two objectives aimed to be optimized are the
niform illumination and the energy efficiency. The proposed sys-
em works then with the default parameters presented throughout
ection 3. The only difference is due to enable the comparison be-
ween the two methods. For this, the only available set of lighting
oles for the proposed system is both 10 m pole height and 100 W
amps (referred in Table 2), the same ones used by the lumen method.
The resulting isolux curve is shown in Fig. 16. Table 4 contains
ome quality measurements to compare both results (lumen method
nd proposed system), even disregarding the obvious advantage of
LCAutoD-EA in respecting geometrical constraints.
Even that both methods could not reach the desired average
f 30 lx, both respect the standards. Also, both achieved energy
lass A. If considering that the lumen method gives the ideal
ompromise between desired average illuminance and electrical
ower consumption, the proposed system uses 22.2% more elec-
rical power to ensure a 17.5% better global uniformity (20.7%
etter, regarding extreme uniformity). Note that those differences
re due to respect the allowed subareas for lighting poles posi-
ioning (part of geometrical constraints). In this way, it is fea-
ible to say that ELCAutoD-EA achieved a good performance
ndeed (Table 6).
It is important to notice that both automated designs (considering
he lumen method with a certain level of automation) can be target
f a specialized engineer discretion.
.2. Irregular shaped areas
To test the efficiency of the ELCAutoD-EA system, a defying area
as elected as a case study. One of the available private parking lots of
he airport of Uberlandia (Brazil) has an irregular layout and has sev-
ral and important subareas where the placement of lighting poles
s restricted. The parking lot of interest has already installed lighting
oles. Fig. 17 presents the area simplified layout. In dark gray is the
utside area (not subject of illumination). In gray, the subareas where
lighting pole can be placed. In light gray, the subareas where light-
ng poles are forbidden (car ways).
The comparison is performed between the current installation
nd the design evolved by the proposed system. The real exterior
ighting design was developed by a specialized and experienced en-
ineer with the aid of available commercial softwares. The installed
esign is shown in Fig. 18. Note that the photometric curve of related
amps was changed to fit the isotropic one used in this research, for
he sake of comparison. The design uses 6 lighting poles (all 12 m
eight), where 4 of them has three 250 W lamps each and the other
wo has two 250 W lamps each.
Again, the proposed system works then with the default parame-
ers presented throughout Section 3. For this case study, pole heights
f 4, 7, 10 and 12 m were available, as well as luminaires whose pos-
ible lamps are enlisted in Table 2. The flowchart of the evolution-
ry algorithm (using NSGA-II) is presented in Fig. 2. After the evolu-
ionary algorithm run, the set of Pareto optimal solutions regarding
oth illumination uniformity and energy efficiency could be found in
ront#1. The winner design to be presented to the user is taken re-
pecting the overall performance metric (presented in Section 3.5.
or this case study, the evolved exterior lighting design is shown
n Fig. 19.
The evolved design uses 18 lighting poles between 10 and 12 m.
ach one has one single lamp of 100 W or 150 W. The complete listing
ith lighting poles information is shown in Table 5.
Table 4 contains some quality measurements to compare both re-
ults (specialized designer and proposed system).
Note that both designs achieved energy class A. Considering the
verage illuminance, the design of the engineer is approximately the
ouble of the standard requirement. Even if arguing that a great illu-
inance could be well received by the parking lot users, one should
emind that the objective of the standard is to ensure a good level of
isualization for a given type of area. In this perspective, the evolved
esign is better to achieve an average closer to the desired. Also, re-
arding energy efficiency and uniformity, the proposed system uses
7.5% less electrical power to ensure a 227.3% better global uniformity
466.7% better, regarding extreme uniformity).
Note that the CAutoD evolved by the proposed system can be tar-
et of a specialized engineer discretion. This is the case mainly when
n need of adjusting some lighting pole placements due to aesthetic
urposes, especially because material availability is an input data for
he proposed system.
220 H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222
Fig. 17. Case study, the simplified blue print of a defying irregular shaped parking lot with 3,550 m2 (160 × 30 m, overall).
Fig. 18. Isolux curve from exterior lighting design performed by the specialized engineer.
Fig. 19. Isolux curve from exterior lighting design evolved by the ELCAutoD-EA system.
c
j
r
m
d
s
4.3. Cluster and speed-up
This section discuss the idea to justify the use of a cluster to han-
dle the problem of computer automated exterior lighting design. As
an NP-hard problem, multi-criteria optimization benefits from paral-
lelism, in special when using evolutionary algorithms through island-
model to achieve supra-linear speed-ups (Collet et al., 2013).
Table 7 presents speed-ups achieved when using the cluster in
ontrast of using a single core machine. As could be seen, speed-up
ustifies the use of the cluster. Note that this table is missing the pa-
ameters actually used in the case study because there is no available
emory to run 20,000 individuals in a single core machine (500 in-
ividuals per island/core used in the case study). Also, the theoretical
peed-up of 40 × is surpassed by experimental runs.
H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222 221
Table 5
Lighting poles listing for the evolved CAutoD.
ID Lamp for Pole Positioning
the luminaire height (m) coordinates
A NAV-E 100W SUPER 4Y 10 (18.0,12.0)
B NAV-E 100W SUPER 4Y 10 (167.0,35.0)
C NAV-E 100W SUPER 4Y 12 (22.0,29.0)
D NAV-E 100W SUPER 4Y 12 (49.0,12.0)
E NAV-E 150W SUPER 4Y 10 (157.0,38.0)
F NAV-E 150W SUPER 4Y 10 (99.0,17.0)
G NAV-E 150W SUPER 4Y 12 (140.0,37.0)
H NAV-E 150W SUPER 4Y 12 (7.0,26.0)
I NAV-E 150W SUPER 4Y 12 (124.0,37.0)
J NAV-E 150W SUPER 4Y 12 (132.0,20.0)
K NAV-E 150W SUPER 4Y 12 (82.0,35.0)
L NAV-E 150W SUPER 4Y 12 (35,35)
M NAV-E 150W SUPER 4Y 12 (69.0,21.0)
N NAV-E 150W SUPER 4Y 12 (55.0,35.0)
O NAV-E 150W SUPER 4Y 12 (10.0,36.0)
P NAV-E 150W SUPER 4Y 12 (114.0,25.0)
Q NAV-E 150W SUPER 4Y 12 (34.0,6.0)
R NAV-E 150W SUPER 4Y 12 (103.0,39.0)
Table 6
Comparison between human-made design and the ELCAutoD-EA system (best values
of each column in bold).
Designed by Electrical power
consumption
(kWh)
Average
illuminance
Uniformity Energy class
U0 Ug
Lighting engineer 4.0 66.64 lx 0.22 0.09 A (0.54)
ELCAutoD-EA 2.5 27.44 lx 0.72 0.51 A (0.78)
Table 7
Speed-ups achieved by using a cluster of computers instead of one core to the com-
puter automated exterior lighting design for a parking lot of the airport of Uberlandia
(Brazil).
1 core 40 cores Speed-up
Individuals/core Exec. time [s] Individuals/core Exec. time [s]
2000 2738.7 50 44.88 ≈ 61 ×3000 4782.4 75 68.10 ≈ 70 ×4000 8057.9 100 98.12 ≈ 82 ×
c
a
t
a
p
e
t
5
p
t
l
s
e
d
l
s
o
q
p
i
p
m
a
p
m
w
v
c
t
u
g
d
i
u
fi
t
t
n
p
s
i
t
c
s
l
t
l
m
m
b
w
i
d
b
N
p
p
s
c
a
h
t
d
t
i
l
t
t
s
s
g
g
t
d
i
i
p
s
b
i
b
As an approximation, i.e. considering enough RAM memory, if the
ase study was supposed to run in a single core machine and adopting
power regression, the execution time for 20,000 individuals would
ake more than 26 h, while with 500 individuals per core the cluster
ctually ends up 70 generations in less than 8 min. As an rough ap-
roximation, a logarithmic regression applied to speed-ups shows an
xpected speed-up of ≈ 128 × for the actual case study if compared
o the single core run.
. Conclusion
This work presents the ELCAutoD-EA system, an application to
erform computer-automated exterior lighting design, aimed to fill
he void of computational intelligence based programs destined to
ighting engineering professionals. This work could be seen as a
uccessful non-trivial expert system. ELCAutoD-EA implements an
volutionary algorithm to perform as an expert system to lighting
esigns. Based on different user inputs, this system presents a mal-
eable structure which allows it to handle different constraints, re-
pect all applicable standards and optimize compromise between
pposite criteria (low power consumption and high illumination
uality).
Besides computer aided design softwares, lighting engineers lack
roper computational intelligence tools to help with the increas-
ng volume of information to be aware in time of designing new
rojects. The ELCAutoD-EA system can help by optimizing hard
ulti-objective problems. From the point of view of evolutionary
lgorithm field, a new chromosome is proposed here, as well as
roper genetic operators, in order to keep the structure of the design
alleable to be initiated, robust against invalid designs and generic
hen accomplishing different user options. Finally, from the point of
iew from lighting engineering, this is the only system that allows
omputer-automated design on irregular shaped areas with restric-
ions on lighting pole positioning. From the point of view from the
ser, the system is very intuitive and decisions to be taken are well
uided.
Also, the use of web clients and cluster of computers turns any
esign demand into an everyday experience, i.e., user can remotely
nput related data, ask for the design and receive it after a few min-
tes without having to stay connected to wait for it. User can de-
ne which materials (e.g. pole heights, lamps, fixtures) he desires
o work with, as well as set up the area layout (defining restric-
ive subareas for lighting pole placements), desired average illumi-
ance (with tolerance), and parallel evolutionary algorithm related
arameters.
A discussion on advantages and disadvantages of the presented
ystem is only possible if the area of interest is regular and every place
nside the area is allowed to position a lighting pole; otherwise, au-
hors do not known any other system or method to be used in such
omparison. A known limitation of this system is to not deal with ob-
tacles, i.e., fixed objects inside an area which forbid positioning of
ighting poles over them and, at the same time, restrict light to pass
hrough them (e.g., a small building, billboards, trees).
With this in mind, a simple example was shown, aiming to il-
uminate a regular shaped area (a square). To compare the perfor-
ance of the proposed system, authors have applied the classic lu-
en method to this area in order to achieve the ideal compromise
etween desired average illuminance and power consumption. After-
ards, some sub-areas were defined in where lighting pole position-
ng is forbidden. The ELCAutoD-EA system automatically achieved a
esign which both uses 22.2% more power than the design obtained
y the lumen method and ensures a 17.5% better global uniformity.
ote that the proposed design respects restrictions on lighting pole
lacements.
As a case study, a real world area (a parking lot from the air-
ort of Uberlandia) was chosen. The area layout is irregular and has
everal restrictions on lighting pole placements. The lumen method
ould not deal with this case, so the lighting facility was designed by
specialized engineer using his experience and some simulators to
elp with the final illumination visualization. The ELCAutoD-EA sys-
em was used to provide an alternative design. When comparing both
esigns, the evolved one achieved an average illuminance closer to
he standards, while the implemented design did not. Also, regard-
ng energy efficiency and uniformity, the evolved design uses 37.5%
ess electrical power to ensure a 227.3% better global uniformity. Bet-
er uniform illumination and energy efficiency were achieved. A bet-
er illumination quality is only possible because the proposed system
olves multi-objective optimization problems which mean that the
ystem avoids energy wastes that are often not clear to a lighting en-
ineer in charge of a given project.
Some subjects are envisioned to be object of future works. Re-
arding the program itself, there are some differences on the way
o implement functionalities to be explored, as other types of non-
ominated sorting schemes to multi-objective optimization. Regard-
ng parallel hardware support, the intention is to plug GPGPU cards
n computers from the cluster and test performance gains with CUDA
rogramming. Regarding the application itself, other functionalities
hould be explored, as the inclusion of fix obstacles within the area to
e illuminated or even to adapt the system to communicate through
nternet with proper hard wares database in order to optimize a given
udget as well.
222 H. Rocha et al. / Expert Systems With Applications 45 (2016) 208–222
K
K
K
L
L
M
M
OP
P
R
R
S
S
Z
Z
References
ABNT (2012). NBR 5101 public lighting. (in Portuguese).
Brownlee, A. E., & Wright, J. A. (2015). Constrained, mixed-integer and multi-objective
optimisation of building designs by NSGA-ii with fitness approximation. AppliedSoft Computing, 33, 114–126. doi:10.1016/j.asoc.2015.04.010.
Carlucci, S., Cattarin, G., Causone, F., & Pagliano, L. (2015). Multi-objective optimizationof a nearly zero-energy building based on thermal and visual discomfort mini-
mization using a non-dominated sorting genetic algorithm (NSGA-ii). Energy andBuildings, 104, 378–394. doi:10.1016/j.enbuild.2015.06.064.
Castro, F., del Acebo, E., & Sbert, M. (2012). Energy-saving light positioning using
heuristic search. Engineering Applications of Artificial Intelligence, 25(3), 566–582.doi:10.1016/j.engappai.2011.11.009.
CELG (2010). Annual administration report 2010. Retrieved from http://celgd.celg.com.br/arquivos/financeiro/2010/CELGD_DEMONSTRACOES_2010.pdf. (in Portuguese).
Accessed 13.07.11.Collet, P., Krüger, F., & Maitre, O. (2013). Automatic parallelization of EC on GPG-
PUs and clusters of GPGPU machines with EASEA and EASEA-CLOUD. In S. Tsut-sui, & P. Collet (Eds.), Massively parallel evolutionary computation on GPGPUs. In
Natural Computing Series (pp. 35–59). Springer Berlin Heidelberg. doi:10.1007/
978-3-642-37959-8_3.Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). A fast and elitist multiobjective
genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2),182–197. doi:10.1109/4235.996017.
European Commission (2013). Energy strategy for Europe. Retrieved from http://ec.europa.eu/energy/index_en.htm. Accessed 13.07.11.
Goldberg, D. E. (1989). Genetic algorithms in search, optimization and machine learning.
Addison-Wesley Longman.Holland, J. H. (1975). Adaptation in natural and artificial systems. The University of Michi-
gan Press.Hornby, G. S., Lohn, J. D., & Linden, D. S. (2011). Computer-automated evolution of an
x-band antenna for NASA’s space technology 5 mission. Evolutionary Computation,19(1), 1–23.
Horváth, I., & Vroom, R. W. (2015). Ubiquitous computer aided design: a broken
promise or a sleeping beauty? Computer-Aided Design, 59, 161–175. doi:10.1016/j.cad.2014.10.006.
IESNA (2000). Lighting handbook reference & application. Illuminating Engineering So-ciety of North America IESNA.
Kamentsky, L., & Liu, C. N. (1963). Computer-automated design of multifont print recog-nition logic. IBM Journal of Research and Development, 7(1), 2–13. doi:10.1147/rd.71.
0002.
amjoo, A., Maheri, A., Dizqah, A. M., & Putrus, G. A. (2015). Multi-objective designunder uncertainties of hybrid renewable energy system using NSGA-ii and chance
constrained programming. International Journal of Electrical Power & Energy Sys-tems, 74, 187–194. doi:10.1016/j.ijepes.2015.07.007.
imovski, D., Ortega, J., Ortiz, A., & Baños, R. (2015). Parallel alternatives for evolution-ary multi-objective optimization in unsupervised feature selection. Expert Systems
with Applications, 42(9), 4239–4252. doi:10.1016/j.eswa.2015.01.061.irby, R. M., & Karniadakis, G. (2003). Parallel scientific computing in C++ and MPI: A
seamless approach to parallel algorithms and their Implementation. Cambridge.
ighting Europe (2013). Eco-lighting project: Preliminary background report. BoulevardA. Reyers 80, 1030 Brussels..Technical Report
uque, G., & Alba, E. (2011). Parallel genetic algorithms theory and real world applications.Berlin Heidelberg: Springer-Verlag.
inisterio de Industria, Turismo y Comercio de España (2008). Reglamento de eficien-cia energética en instalaciones de alumbrado exterior y sus instrucciones técnicas
complementarias ea-01 a ea-07. (in Spanish).
PICH (2012). Implementation of MPICH2. Retrieved from http://www.mcs.anl.gov/research/projects/mpich2/. Accessed 13.07.11.
SRAM (2013). Luminotechnical practical manual. (in Portuguese).hilips LiDAC (2011). CalcuLuX 7.6.1 - area. Retrieved from http://www.lighting.philips.
com. Accessed 13.07.11.ROCEL (2004). Procel reluz : Instruction manual. (in Portuguese).
abaza, O., Pena-Garcia, A., Perez-Ocon, F., & Gomez-Lorente, D. (2013). A simple
method for designing efficient public lighting, based on new parameter relation-ships. Expert Systems with Applications, 40.
out, P. K., & Acharya, D. P. (2015). Fast physical design of CMOS ROs for optimal per-formance using constrained NSGA-ii. AEU - International Journal of Electronics and
Communications, 69(9), 1233–1242. doi:10.1016/j.aeue.2015.05.004.haygan, M., Alimohammadi, A., Mansourian, A., Govara, Z. S., & Kalami, S. M. (2014).
Spatial multi-objective optimization approach for land use allocation using NSGA-
ii. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing,7(3), 906–916.
henfield, A., & Fleming, P. J. (2014). Multi-objective evolutionary design of robustcontrollers on the grid. Engineering Applications of Artificial Intelligence, 27, 17–27.
doi:10.1016/j.engappai.2013.09.015.avala, G. R., Nebro, A. J., Luna, F., & Coello Coello, C. A. (2014). A survey of multi-
objective metaheuristics applied to structural optimization. Structural and Multi-
disciplinary Optimization, 49(4), 537–558. doi:10.1007/s00158-013-0996-4.hou, A., Qu, B.-Y., Li, H., Zhao, S.-Z., Suganthan, P. N., & Zhang, Q. (2011). Multiobjective
evolutionary algorithms: A survey of the state of the art. Swarm and EvolutionaryComputation, 1(1), 32–49. doi:10.1016/j.swevo.2011.03.001.
