Building and Environment 46 (2011) 331e338

Contents lists available at ScienceDirect

Building and Environment

journal homepage: www.elsevier.com/locate/buildenv

Multi-objective optimization as a new approach to illumination design of

interior spaces
Fabiano Cassol a, Paulo Smith Schneider a, Francis H.R. França a, *, Antônio J. Silva Neto b
a
Department of Mechanical Engineering, Federal University of Rio Grande do Sul, Porto Alegre, RS, Brazil
b
Polytechnic Institute, IPRJ, University of the State of Rio de Janeiro, Nova Friburgo, RJ, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: The illumination design is in general oriented towards attaining a prescribed illuminance on a work
Received 26 April 2010 plane with the most economical lighting configuration. This work presents a new methodology to find
Received in revised form the location and the luminous power of the light sources to satisfy a prescribed illuminance on the work
25 July 2010
plane with the constraint of lowest power consumption. Mathematically, the illumination design is
Accepted 26 July 2010
inherently an inverse problem in which the boundary conditions associated to the light sources are
unknown. In the proposed methodology, the inverse analysis is formulated as an optimization problem,
Keywords:
in which the objective function aims at minimizing the total power of the light sources as well as the
Illumination design
Multi-objective optimization
deviation between the prescribed and the resulting illuminance. The optimization problem is solved with
Inverse analysis the generalized extremal optimization (GEO) algorithm. The lighting calculations are based on the
Radiosity method radiosity method. The examples discussed in the paper consider the prescription of uniform illuminance
at the bottom surface of a rectangular space. The methodology leads to a set of design solutions that are
capable of satisfying the problem with different levels of accuracy and energy consumption.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction calculation and photorealistic architectural rendering [2e10]. With

Although illumination has long been an inherent part of archi-
tectural design, only more recently the art of illuminating spaces
has evolved to become a science. The evolution arose from the
gained understanding of the nature of light and the establishment
of the equations that describe the transport of light in spaces. Due
to the complexity of the light transport, the first methods involved
various approximations to allow rapid, simple estimations of the
illumination of a space from daylight and/or artificial light. Of the
first approaches, the lumen method [1] still finds widespread use
among illumination designers, but cannot account for many
complexities (such as shading, non-uniform distribution of the
luminaries in the room, etc.), and at best can satisfy the specified
illumination only in the average sense. Other more realistic
methods, such as ray tracing, radiosity or hybrid approaches, arose
from techniques that were developed by the computer graphics
and the radiative heat transfer communities. Those methods can
account for the transport of light in fine detail, but on the other
hand require computational tools for the solution of the resulting
equations. In the last few decades, a number of free and commercial
codes have been developed to assist in engineering lighting
* Corresponding author. Tel.: þ55 51 3308 3360; fax: þ55 51 3308 3222.
E-mail address: frfranca@mecanica.ufrgs.br (F.H.R. França).
the computational techniques, it is possible to perform illumination
calculations considering specular, diffuse and directional-diffuse
reflection, combined daylight and artificial illumination and
complex geometry settings.
In spite of the advances in lighting calculations, the designer can
still face considerable difficulty to design an illumination system that
satisfies with accuracy the desired illuminance on the work plane.
For instance, several applications, including manufacturing
processes, plants growth and domestic animals reproduction,
require specific illumination intensity [1]; another problem of
interest is controlling the luminous environment to meet human
satisfaction [11e14]. Employing the conventional techniques, the
designer must specify the location and power of the light sources,
and run a suitable computational code to find the resulting illumi-
nance on the work plane. If the achieved illuminance is not the one
required, it is necessary to redo the calculation until a satisfactory
solution is achieved. Due to the complexity of the light transport, it is
in general very difficult to make a new guess from a previous
attempt. The designer is left to choose the best solution from
a collection of trials, which is probably not the best possible one.
A more effective approach would be a direct determination of the
location and power of the light sources from the specified illuminance.
A similar problem, in both terms of mathematical formulation and
specification of the boundary conditions, is the thermal design of

0360-1323/$ e see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.buildenv.2010.07.028
332 F. Cassol et al. / Building and Environment 46 (2011) 331e338

purely radiative systems, in which the power of the heaters is to be
determined to provide the specified heat flux on the material under-
going a thermal process. Solutions of this type of problem appear in
[15e18], applying a technique that is known as inverse design.
Although inverse design is capable of avoiding the trial-and-error
procedure of the conventional approach, it invariably leads to ill-
conditioned systems of equation, which can be in addition over or
under-determined. Therefore, the inverse formulation of radiative
systems cannot be solved by standard matrix solvers, but require
regularization methods or optimization techniques. The application of
inverse design to illumination appears in [19]. The authors applied two
approaches, matrix regularization and optimization, to determine the
luminous power of the light sources to attain uniformity of illumi-
nance at the bottom surface of a rectangular space. As typical in inverse
design, the problem did not have an exact solution that was physically
meaningful, but different approximate solutions were found to satisfy
the specified illuminance within some level of accuracy.
A common aspect of the inverse solutions presented in Refs.
[15e19] is that the heaters and the light sources were arranged in
a certain spatial configuration, so the optimum solutions were
constrained by the chosen configuration, and so did not represent
a global optimization. As an advance to the previous studies, the
present work applies the inverse analysis to find the luminous
power and the location of the light sources to attain the specified
illuminance on the work plane. This work also discusses how the
optimization technique can be formulated to lead to simultaneous
minimization of two quantities: (i) the deviation between the
specified and the actual illuminance (accuracy), and (ii) the amount
of luminous power that is required to satisfy the desired illumi-
nation (energy efficiency). The proposed methodology is applied to
on the work plane, light sources and walls in the one-quarter of the
system are designated by KP, KL and KW, so 1  kp  KP, 1  kl  KL
and 1  kw  KW. When a general relation applies to any kind of
surface element, the general index k will be used.
For illustration of the proposed methodology, it is considered that
a uniform illuminance, indicated by Li,specified (in units of lumens/m2
or lux), is specified on the work plane. The design problem consists
of finding the location and power of the light sources to attain the
specified illuminance. The incident illuminance on element kp in the
work plane, Li,kp, in lux, is a result of multiple reflections of the light
in all surfaces, and can be computed from:
X
KP X
KL X
KW
Li;kp ¼ Fkpkp* Lo;kp* þ Fkpkl Lo;kl þ Fkpkw Lo;kw (1)
kp*¼1 kl¼1 kw¼1
where Lo,k (in lux) is the outgoing illuminance or luminous
radiosity of surface element k, which takes into account both
emission and reflection, and Fkk* is the view factor between
surface elements k and k*. When elements k and k* are on the same
plane, the respective view factor will be equal to zero, as with the
work plane elements in Fig. 1(a). In spite of this, all view factors are
being kept in the formulation to accommodate the cases in which
the elements are not all in the same plane.
For a more general presentation of the solution, all results in this
paper are presented in dimensionless form. The dimensionless inci-
dent illuminance on each work plane element is li,kp ¼ Li,kp/Li,specified,
so the target of the work plane is that li,kp be as close as possible of the
unit. Eq. (1) can be expressed in dimensionless form by:
X
KP X
KL X
KW

a rectangular cavity formed with perfectly diffuse walls, in which all li;kp ¼ Fkpkp* lo;kp* þ Fkpkl lo;kl þ Fkpkw lo;kw (2)
kp*¼1 kl¼1 kw¼1
the illumination is provided by the light sources. The lighting
calculations are solved with the radiosity method. The inverse
where lo,kp ¼ Lo,kp/Li,specified is the dimensionless outgoing illumi-
problem is formulated as an optimization problem, and is solved
nance of surface element k. According to Eq. (2), to determine the
with the generalized extremal optimization (GEO) algorithm [20].
incident illuminance on the work plane, it is first necessary to
determine the dimensionless outgoing illuminance of all surface
2. Physical and mathematical modeling elements, which are given by the following relations:
0
A schematic representation of the system is shown in Fig. 1(a). X
KP X
KL
The three-dimensional rectangular enclosure is formed by surfaces lo;kp ¼ rkp @ Fkpkp* lo;kp* þ Fkpkl lo;kl
that are perfectly diffuse. The work plane, where the incident kp* ¼ 1 kl ¼ 1
1 (3)
illuminance is to be specified, is located on the bottom surface of X
KW
the enclosure; the light sources are located on the top surface. The þ Fkpkw lo;kw A
remaining of the enclosure is formed by walls that partially reflect kw ¼ 1
incident light. The length, width and height of the enclosure are 0
designated by L, W and H, respectively. Due to the physical X
KP X
KL
symmetry of the examples considered in this paper, only one- lo;kl ¼ ekl þ rkl @ Fklkp lo;kp þ Fklkl* lo;kl*
quarter of the system needs to be solved, as depicted in Fig. 1(b). kp ¼ 1 kl* ¼ 1
The enclosure is divided into finite-sized square elements,
1
X
KW
Dx ¼ Dy ¼ Dz, in which the luminous energy balance will be þ Fklkw lo;kw A (4)
applied. The work plane, light sources and walls elements are kw ¼ 1
designated by kp, kl and kw, respectively. The number of elements

light
sources

Wp work plane
H
W Lp
y z
L x

Fig. 1. (a) Representation of the illumination system; (b) Physical domain: one-quarter division of the illumination system due to symmetry.
 F. Cassol et al. / Building and Environment 46 (2011) 331e338
Fig. 2. Design variables encoded in a binary string.
0
X
KP X
KL
lo;kw ¼ rkw @ Fkwkp lo;kp þ Fkwkl lo;kl
kp ¼ 1 kl ¼ 1
1 locations of the light sources were known. In this work, the location
X
KW
þ Fkwkw* lo;kw* A
kw* ¼ 1
In the above equations, rk is the hemispherical reflectivity of surface
element k in the visible range of the spectrum. The dimensionless
luminous power of the light source element kl is defined as ekl ¼ Ekl/
Li,specified, where Ekl (in lux) is its luminous power. As described by
Eqs. (3)e(5), the outgoing illuminance of the work plane and wall
elements correspond solely to the reflection of the incident illumi-
nance; no light is emitted from the work plane and wall elements,
so the luminous powers are null for those elements (ekp ¼ ekw ¼ 0).
For the light sources elements, the outgoing illuminance is the sum of
the luminous power and the reflection of the incident illuminance.
In the conventional forward design of illumination systems, the
dimensionless power and location of each light source element, ekl,
is prescribed. It follows that Eqs. (3)e(5) form a system of linear
equations on the outgoing illuminances of all surface elements. The
system presents the same numbers of equations (i.e., one equation
for each element) and of unknowns (i.e., the outgoing illuminance of
each element), and can be solved by any standard matrix solver, such
as GausseSeidel and Gaussian elimination. After the solution of the
lo,k’s, Eq. (2) is applied to each work plane element kp to determine
its dimensionless illuminance, li,kp, which is then compared to the
desired value, li,specified ¼ 1. The drawback of this approach is that
choosing the location and luminous power of the light sources to
lead to the specified illuminance on the work plane is very difficult.
In the conventional approach, a number of solutions are tried, and
then the best attempt of the set is selected. However, the selected
solution will probably not be the best possible solution.
The inverse design methodology described in Refs. [15e19]
proposes the direct inversion of the system formed by Eqs. (3)e
(5), imposing that li,kp ¼ 1 and letting the ekl’s to be the unknowns.
This leads to a system of equations that is ill-conditioned, since it is
a discrete form of a variation of the Fredholm integral of the first kind
Fig. 3. (a) The work plane (shaded area) in one-quarter of the bottom and top surfaces: KP ¼ 108. (b) Light sources (black dots) in one-quarter of the domain: KL ¼ 9. The two
surfaces are divided into 15 and 12 equal-sized elements in the x and y directions, respectively.
333
[21]. In addition, the number of equations and the number of
unknowns are not the same, unless KP ¼ KL. As a consequence, the
solution of the system of equations can render physically meaningful
answers only if regularization methods are applied. One serious
limitation of the direct inversion is that it requires the specification
of the location of light sources, since there seems to be no clear way
to also set their location as unknowns to be determined. Therefore,
the obtained results can be among the best for the proposed
configuration, but not for all possible configurations.
An alternative approach is to treat the inverse problem as an
optimization problem, in which the configurations of the light
sources are sought to minimize the deviation between the specified
and the resulting illuminances on the work plane. This approach
was successfully applied in Ref. [19], but considering that the
(5) and power of the light sources are to be determined. This will be
accomplished by a multi-objective optimization that is guided
towards the accuracy of the solution and the energy efficiency. This
can be accomplished by requiring that the configurations of the
light sources minimize the following multi-objective function:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u KP  2
uX X
KL
F ¼ t li;kp  1 þ a ekl (6)
kp ¼ 1 kl ¼ 1
where KP and KL are respectively the number of work plane
elements and light sources in the one-quarter of the system, Fig.1(b).
According to Eq. (6), the objective function is composed of two
terms. The first term is a measure of the least-square of the devia-
tions between the specified illuminance (which, in dimensionless
form, it is equal to the unity) and the illuminance on each element kp
that is obtained from a given configuration of the light source. The
second term is a measure of the total luminous power that should be
provided by the light sources. Parameter a establishes how much the
optimization will lean towards the accuracy of the solution or
towards the energy efficiency. When a is zero, the optimization will
lead to a configuration of the light sources that provides the most
accurate solution; when a is increased, the search for the solution
will not only be guided by accuracy of the illumination on the work
plane, but also by the minimization of the total luminous power
required in the light sources. For a meaningful balance between
these two aspects of the design, accuracy and energy efficiency, it is
necessary to find appropriate values for a. A simple procedure to find
these values is described in Section 4, Results and Discussion.
In the minimization of the objective function, most of the
optimization approaches have been based on gradient based
methods [22], but such methods, when convergence occurs, may in
fact lead to local minima instead of global minimum. Recent years
have seen an increasing interest towards stochastic global search
334 F. Cassol et al. / Building and Environment 46 (2011) 331e338

Fig. 4. (a) Objective function F as a function of parameters s and a, corresponding to the lowest value from fifty independent runs, each one with ten thousand evaluations.
(b) Objective function F as a function of the number of evaluations for s ¼ 1.25. Case 1.

methods [23e27]. Even though such methods are usually compu-
tationally intensive, when properly implemented they may lead to
the vicinity of the global minimum. Recently, a new stochastic
optimization algorithm inspired by a simplified evolutionary model
was proposed, the so called generalized extremal optimization
(GEO) [20]. This algorithm was developed to be easily applicable to
a broad class of nonlinear constrained optimization problems,
while having only one parameter to be adjusted. The GEO algorithm
has already been applied in the solution of inverse radiative
transfer problems [28,29], and also to illumination design [19], and
is described in the next section.
3. The generalized extremal optimization (GEO) algorithm
The GEO algorithm was devised to improve the extremal opti-
mization method [30] for general application to optimization
problems. In the GEO algorithm, the M design variables of the
optimization problem, bm, are encoded in a string that is similar to
a chromosome in a genetic algorithm (GA) with binary represen-
tation, as seen in Fig. 2. The string is formed by N bits (species) that
can assume the values of 0 or 1. The species are aligned and
assigned a fitness value that will determine the species that are
more prone to mutate. The fitness number is proportional to the
gain (or loss) that the objective function value presents when
the bit is mutated (flipped). All bits are then ranked from r ¼ 1, for
the least adapted bit, to r ¼ N for the best adapted. A bit is then
mutated (flipped) according to the probability distribution below:
a given stopping criterion is reached. In this work, the number of
evaluations of the objective function, G, is chosen to stop the
algorithm. A discussion on how the parameters s and G can be
selected appears in Section 5, Results and Discussion.
In the GEO algorithm, the definition of the number of bits that
will represent each design variable is made by simply setting for
each variable the number of bits necessary to assure a desirable
precision. For continuous variables the minimum number of bits
that are necessary to achieve a certain precision can be obtained
from the following relation [20]:
  
bm;u  bm;l
2nm  þ1 (8)
pm
where nm is the minimum number of bits associated to variable bm,
bm,u and bm,l are the upper and lower bounds of the variable, and pm
is the desired precision. In practice, nm is taken as the smallest
integer that satisfies Eq. (8). The total number of bits N is then given
by the summation of all nm’s. The physical value of each design
variable is obtained through the equation:
  Im
bm ¼ bm;l þ bm;u  bm;l   (9)
2nm  1
in which Im is the integer number obtained in the transformation of
the variable bm from its binary form to a decimal representation.
4. Solution procedure

PðrÞ ¼ r s (7) The illumination design was formulated as an optimization

problem to minimize the objective function F defined by Eq. (6).
in which (1  r N). In Eq. (7), s is a positive adjustable parameter:
The optimization involves different kinds of design variables, that
for s / 0, the algorithm becomes a random walk, whereas for
is, the locations and the power inputs of the light sources. The
s / N, the search is deterministic. This process is repeated until
objective function can be expressed by the following functional:

Table 1
F ¼ Fðb1 ; .; bm ; .; bM Þ (10)
Required dimensionless luminous power of the light sources based on the mini-
The M design variables b1 to bM represent the location and the
mization of the objective function F, Eq. (6). Case 1: light sources having fixed
location, but unknown luminous power. KL ¼ 9. power of the light sources. The locations themselves are repre-
sented by coordinates xm and ym, or alternatively by the grid indices
a¼0 a ¼ a* a ¼ 5a*
im and jm. The optimization problem is subjected to the following
kl ikl jkl ekl ikl jkl ekl ikl jkl ekl constraints:
1 3 3 38.3 3 3 36.7 3 3 19.0
2 3 7 20.8 3 7 20.0 3 7 38.0 il  im  iu ðx direction restrictionÞ (11a)
3 3 11 25.6 3 11 24.0 3 11 17.6
4 8 3 25.5 8 3 25.6 8 3 44.8
5 8 7 19.0 8 7 22.4 8 7 3.9 jl  jm  ju ðy direction restrictionÞ (11b)
6 8 11 12.7 8 11 13.0 8 11 19.5
7 13 3 29.2 13 3 33.1 13 3 20.8
el  em  eu ðluminous power restrictionÞ (11c)
8 13 7 18.5 13 7 12.7 13 7 25.6
9 13 11 17.6 13 11 19.0 13 11 14.3 To minimize the objective defined by Eq. (6), the following
PKL PKL PKL
kl ¼ 1 ekl ¼ 207:2 kl ¼ 1 ekl ¼ 206:5 kl ¼ 1 ekl ¼ 203:5
procedure is followed:
 F. Cassol et al. / Building and Environment 46 (2011) 331e338 335

Fig. 5. Dimensionless illuminance on the work plane for Case 1. Optimization with light sources having different luminous powers: (a) a ¼ 0; (b) a ¼ a*; and (c) a ¼ 5a*. (d) The light
sources are constrained to have the same luminous power, el ¼ 21.1, which was found by direct sweeping of el in the interval 0  el  50 to minimize F with a ¼ 0.

1. Define the required precision pm for each variable to find the 5. Run the GEO algorithm to select and test other sets of location
respective number of bits, nm, using Eq. (8); and luminous power of the light sources, solving for the inci-
2. Choose the two parameters that govern the GEO algorithm: s in dent illuminance on the work plane elements and determining
Eq. (7) and the number of evaluations of the objective function, the objective function for each set;
G (see discussion in Section 5, Case 1); 6. Repeat the process until the number of evaluations of the
3. Start with a randomly chosen set of location and luminous objective function is reached.
power of the light sources;
4. Solve the system of equations described in Section 2 to find the The verification of the inverse solution is rather straightforward.
incident illuminance on the work plane elements, li,kp, For each configuration of the light sources, the dimensionless
computing the objective function from Eq. (6); incident illuminances on the work plane elements, li,jp’s, are
determined from the solution of the set of Eqs. (2)e(5), which can
be readily compared to the specified illuminance, li,specified ¼ 1. The
average error of the design is given by:
Table 2
KP
Required dimensionless luminous power of the light sources based on the mini-
1 X
li;kp  1  100%
mization of the objective function F, Eq. (6). Case 2: light sources having unknown gavg ¼ (12)
location and equal-valued luminous power. KL ¼ 9. KP
kp ¼ 1
a¼0 a ¼ a* a ¼ 5a*
The maximum error of the design is defined by:
kl ikl jkl ekl ikl jkl ekl ikl jkl ekl

1 3 3 22.8 2 7 22.7 3 5 22.2 gmax ¼ max li;kp  1  100% (13)
2 4 4 22.8 3 4 22.7 3 6 22.2 kp
3 4 8 22.8 3 11 22.7 4 11 22.2
4 4 11 22.8 7 2 22.7 6 3 22.2
5 8 4 22.8 7 4 22.7 8 3 22.2
6 10 10 22.8 8 10 22.7 8 10 22.2
7 11 4 22.8 12 6 22.7 12 6 22.2 5. Results and discussion
8 14 9 22.8 13 3 22.7 13 10 22.2
9 15 1 22.8 13 10 22.7 14 3 22.2
PKL PKL PKL The test cases considered in this work consider a three-
kl ¼ 1 ekl ¼ 205:2 kl ¼ 1 ekl ¼ 204:3 kl ¼ 1 ekl ¼ 199:8
dimensional enclosure as shown in the schematic representation in
336 F. Cassol et al. / Building and Environment 46 (2011) 331e338

Fig. 6. Location of the light sources (black dots) on the top surface for Case 2: (a) a ¼ 0; (b) a ¼ a*; and (c) a ¼ 5a*. The shaded area represents the work plane on the bottom surface.

Fig. 1(a). The aspect ratio of the enclosure base is W/L ¼ 0.8; the to the closest integer. For instance, if im ¼ 1.4, the index will be set
dimensionless height is H/L ¼ 0.2. The selection of the other as im ¼ 1; if im ¼ 15.9, it will be set as im ¼ 15. This way, index im will
dimensions of the enclosure will require a few considerations. First, vary from 1 to 15, covering all fifteen elements in the x-direction
the work plane should not cover the entire extension of the base, that are depicted in Fig. 3(a). The same procedure was applied to
since the portions close to the corners would be mainly affected by index jm, which are rounded down from 1 to 12.99 to cover the 12
the reflections from the side walls, not from the luminous radiation elements in the y direction. Another aspect of the solution is that
from the light source elements on the top surface. Therefore, the distinct light sources are a priori allowed to occupy the same
work plane dimensions are taken as Lp/L ¼ 0.8 and Wp/L ¼ 0.6. The location in the grid. However, this will not be probably a good
hemispherical reflectivities in the visible light region of the work design configuration, and in fact none of the optimum solutions
plane, of the light sources and of the walls are rkp ¼ 0.1, rkl ¼ 0.1 and obtained in this work led to a configuration where more than one
rkw ¼ 0.5, respectively. The specified dimensions of the enclosure light source occupied the same location.
and the reflectivities are design settings that have an important To illustrate the application of the proposed methodology, it is
effect on the numerical results but not on the way the methodology considered that the number of light sources in each quarter of the
works. The mesh for numerical treatment of the problem involved system is nine, that is, KL ¼ 9. Other numbers of light sources can
fifteen, twelve and four equal-sized elements in the x, y and z also be tested without any change in the proposed methodology. It
directions in the one-quarter of the system, Fig. 1(b). Increasing the is expected that using a larger number of light sources will help
resolution of the mesh did not result in perceptible changes in the improving the uniformity on the work plane, although more light
results. sources will probably increase the cost of the design.
Fig. 4(a) and (b) shows the discretization of the bottom and top As seen in Fig. 3(a), the number of work plane elements in each
surfaces, respectively, into fifteen and twelve elements in the x and quarter of the system is KP ¼ 108. Three different designs are
y directions. The shaded area represents the work surface at the considered. In Case 1, the location of the light sources is fixed and
bottom surface, while the dots represent the elements where the the values of their luminous power are left to vary, leading to nine
light sources are located at the top surface. Due to the symmetry in design variables (that is, the luminous power of the nine light
the problem, indicated by the dashed lines, only one-quarter of the sources). In Case 2, the location and the luminous power of the light
system is solved (0  x/L  0.5, 0  y/L  0.4). The location of each sources are unknown, but their luminous power are imposed to be
light source can be specified by varying the integer indices im and jm all the same, which leads to nineteen variables (the eighteen
in the intervals [il ¼ 1, iu ¼ 16] and [jl ¼ 1, ju ¼ 13]. For the dimen- location indices i and j and the unknown luminous power). Finally,
sionless luminous power of the light sources, em, the chosen in Case 3, the location and luminous power of the light sources, so
interval is [el ¼ 0, eu ¼ 50]. The precision pm was chosen to be 0.1 for that there are twenty-seven variables (the eighteen location indices
all variables. i and j and the nine different luminous powers).
At this point, it is necessary an explanation on how the location For Case 1, in which the location of the nine light sources are
of the light sources were set. When computing im, the results will arbitrarily fixed, the proposed configuration is shown in Fig. 3(b).
vary from 1 to 15.99, but the values, computed from Eq. (9), will not As seen, the nine light sources are regularly distributed on the top
be necessarily integers. In this case, the numbers are rounded down surface. The design aims at finding the power of the light sources
that minimizes the objective function given by Eq. (6). To run
the GEO algorithm, two parameters are needed: s, which affects the
Table 3 probability of mutation (flip) of a bit according to Eq. (7), and G,
Required dimensionless luminous power of the light sources based on the mini- the number of evaluations of the objective function. It is first
mization of the objective function F, Eq. (6). Case 3: light sources having unknown
considered the case in which a ¼ 0, so that only the first term of the
location and luminous power. KL ¼ 9.
objective function is minimized. The GEO algorithm was run 10
a¼0 a ¼ a* a ¼ 5a* thousand times for the evaluation of the objective function F, using
kl ikl jkl ekl ikl jkl ekl ikl jkl ekl different values of s. The procedure was repeated fifty times,
1 1 8 50.0 3 3 32.7 2 9 21.4 starting each time with a new randomly chosen configuration of
2 4 3 31.9 4 9 38.8 2 11 13.9 the light sources. The solution that led to the minimum F was
3 5 11 7.1 4 11 8.7 4 4 25.3 selected as the optimum solution for the given s. The results are
4 7 11 13.9 7 4 37.0 4 5 16.3
shown in Fig. 4(a), under the label a ¼ 0. Inspection of the results
5 8 1 36.8 10 10 18.4 8 2 7.2
6 8 6 2.4 12 2 12.7 8 4 25.7 suggests that the optimum value of s is 1.25. For the next runs, more
7 10 8 26.4 13 3 19.1 8 10 30.0 evaluations were allowed for the algorithm to search the design
8 13 3 33.2 13 6 15.3 13 5 41.7 space. The evolution of the solution towards the minimization of F,
9 14
PKL
10 19.2 13
PKL
10 19.5 14
PKL
11 15.9 can be seen in Fig. 4(b) as a function of the number of evaluations G.
kl ¼ 1 ekl ¼ 220:9 kl ¼ 1 ekl ¼ 202:2 kl ¼ 1 ekl ¼ 197:4
As expected, the objective function decreased continuously with G.
 F. Cassol et al. / Building and Environment 46 (2011) 331e338 337

Fig. 7. Location of the light sources (black dots) on the top surface for Case 3: (a) a ¼ 0; (b) a ¼ a*; and (c) a ¼ 5a*. The shaded area represents the work plane on the bottom surface.

The decrease was more significant for the first 100 thousand
evaluations, and further evaluations did not lead to considerable
decrease in the objective function. Results for the dimensionless
luminous power of the nine light sources are presented in Table 1,
under label a ¼ 0. The next step is to include in the analysis the
second term of the objective function, related to the total power
required in the light sources. For this, it is first necessary to deter-
mine meaningful values for a. This can be accomplished by the
following procedure. Once the luminous powers are computed for
a ¼ 0, an estimation of a reference value for a can be made by
simply taking the ratio between the first and the second terms that
form the objective function in Eq. (6), that is:
0vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
u KP  , !
uX 2 X
KL
a ¼ @t
*
li;kp  1 A ekl
kp ¼ 1 a¼0 kl ¼ 1 a¼0
In the above equation, the index a ¼ 0 was included to indicate that
the values of the two terms that were obtained from the optimi-
zation with a ¼ 0. For Case 1, it was found a* y 0.005.
Choosing a  a* will lead to an optimization leaning towards
the minimization of the deviation between the specified and the
obtained illuminance on the work plane, while a [ a* will force
the optimization towards the minimization of the total power of
the light sources. In the present analysis, which aims at simulta-
neous minimization of both aspects of the design, it should be
taken values such that awa*. Therefore, in addition to a ¼ 0, two
other solutions are considered, a ¼ a* and a ¼ 5a*. Fig. 4(a) shows
that the optimum s did not vary with the different values of a, and
was kept equal to s ¼ 1.25 for the different values of a. The decrease
of the objective function F with the total number of evaluations for
the different choices of a is shown in Fig. 4(b). Table 1 shows the
dimensionless luminous power for the light sources for the three
values of a. As seen, the change in a leads to considerable changes
in the values of the luminous power of the light sources.
Fig. 5(a)e(c) shows the dimensionless incident illuminance on
the work plane for the solutions with a ¼ 0, a ¼ a* and a ¼ 5a* that
are shown in Table 1. As discussed in Section 2, the dimensionless
incident illuminance on each work plane element is determined
from Eq. (2), where the radiosities are determined from the solu-
tion of the system of equations formed by Eqs. (3)e(5). As seen in
the figures, the dimensionless illuminance remains close to the
specified value of the unity for the three solutions, especially when
solution was found by simply sweeping ekl from 0 to 50 and then
selecting the value of ekl that leads to the minimum value of the
objective function F. This procedure led to ekl ¼ 21.1. As seen in
the figure, the error of the design became considerably larger, with
the maximum error reaching 20%. This illustrates the advantageous
of the proposed solution, which allows the luminous powers to
vary among the light sources to find more accurate answers.
Cases 2 and 3 are treated next. In these two cases, the location of
the light sources was not previously specified, as with Case 1, but
was determined in order to minimize the objective function in Eq.
(6). For Case 2, where all the light sources are supposed to have the
same luminous power (although unknown), the optimum value for
s was again equal to s ¼ 1.25. As with Case 1, the problem was
first solved by setting a ¼ 0, then finding a* y 0.006 from Eq. (14).
(14) Table 2 shows the solutions for Case 2 considering a ¼ 0, a* and 5a*.
Table 2 shows that, while the luminous powers did not change
significantly in the three cases, the location of the light sources
varies greatly. This can be better visualized in Fig. 6(a)e(c), which
shows the locations of the light sources for the different values of a.
For case 3, where the power of the light sources is also left to vary
from each other, the results are shown in Table 3. The optimum
value of s was slightly different from the other previous cases, with
s ¼ 1.00 providing the minimum value for the objective function. As
with Case 2, it was found that the reference value of the optimi-
zation parameter was a* y 0.006. Table 3 shows that both the
location and the power of the light sources are very sensitive to the
choice of a. Fig. 7(a)e(c) shows how the optimum location of the
light sources vary with a.
Further comparisons between the solutions for the three cases
are shown in Table 4, which show the behaviors of the objective
function, of the average and maximum errors of the design solu-
tion, and of the total power of the light sources. In general, the
increase in the value of a led to a reduction in the total dimen-
sionless power required in the lamps, imposing however a larger
error of the design solution. It is interesting to note that the solution
Table 4
Values of the objective function F, average error, maximum error and total luminous
power of the light sources for the different design cases and optimization
parameters.
PKL
F gavg (%) gmax (%) kl ¼ 1 ekl

the optimization was solely guided towards the minimization of the Case 1 a¼0 0.1086 0.86 3.07 207.2
deviation between the specified and the resulting illuminance on a ¼ a* 0.2185 0.91 2.74 206.5
a ¼ 5a* 0.6782 1.24 6.55 203.5
the work plane (a ¼ 0). In this case, the average and the maximum
errors, defined by Eqs. (12) and (13), were 0.86% and 3.07%. When Case 2 a¼0 0.1246 0.95 3.85 205.2
the minimization of the total power in the light sources gained a ¼ a* 0.2643 1.01 5.12 204.3
a ¼ 5a* 0.7513 1.15 3.85 199.8
more weight with a ¼ 5a*, the average and the maximum errors
increased to 1.24% and 6.55%. Fig. 5(d) shows the illuminance on the Case 3 a¼0 0.1263 0.98 3.35 220.9
a ¼ a* 0.2719 1.21 4.16 202.2
work plane when the luminous power of the light sources were
a ¼ 5a* 0.7848 1.45 6.51 197.4
constrained to have the same value, and setting a ¼ 0. The optimum
338 F. Cassol et al. / Building and Environment 46 (2011) 331e338
for Case 3 did not lead to the solution with the smallest errors. This
seems to be a contradiction to the fact that the searching spaces of
Cases 1 and 2 are contained in the searching space of Case 3, and so
the latter should provide a solution at least as satisfactory as the
other two cases. The reason for this seemly contradiction is that
Case 3 involves a larger number of variables (twenty seven in total)
than the other cases, and so it should require a larger number of
evaluations. In all cases shown in Table 4, the results were obtained
from the selection of the best result of fifty independent runs of the
GEO algorithm, each one involving 300 hundred thousand evalu-
ations of the objective functions. Employing an AMD Opteron 2210
Dual Core, 1.8Ghz, 8Gb RAM computer, the computational time
required to run each evaluation was of about 0.022 s, so a solution
involving 100 hundred thousand iterations required about 40 min.
6. Conclusions
This paper considered the inverse design of illumination
systems in which the location and luminous power of the light
sources were to be determined to attain a prescribed illuminance
on the work plane. The inverse problem was formulated as an
optimization problem involving therefore two different kinds of
variables: location and luminous power. The proposed multi-
objective formulation allowed simultaneous optimization of the
accuracy of the illuminance on the work plane and of the total
luminous power required in the light sources. However, the design
methodology is not limited to the objective function that was
considered in this paper, and can be applied to other objective
functions, depending on what is intended to be optimized. The
generalized extremal optimization (GEO) algorithm was the
method selected to search the optimum configurations of the light
sources. Despite requiring a large computational effort, as typical of
stochastic methods, it proved robust and flexible to find different
solutions under different design constraints. Another important
advantage is that the method easily accommodated the task to find
the locations of the light sources, which would be quite difficult to
achieve with direct inversion and regularization of the governing
equations. The examples shown in the paper demonstrated that the
application of multi-objective optimization to illumination design
can lead to accurate solutions for the specified illumination on the
work plane and, at the same time, incorporate the efficiency of the
illumination system in the analysis.
Acknowledgements
FC thanks CNPq due to its financial support by means of
a Masters Degree scholarship. FHRF and AJSN thank CNPq for
research grants 304535/2007-9 and 300171/97-8, respectively.
AJSN also thanks FAPERJ.
References
[1] Rea MS. The IESNA lighting handbook. New York: Illuminating Engineering
Society of North America; 2000.
[2] ADELINE 3.0. SCH task 12. International Energy Agency, FraunhofereInstitut
fur Bauphisik; 2002.
[3] Bourgeois D, Reinhart CF, Ward G. A standard daylight coefficient model for
dynamic daylighting simulation. DDS. Canada: National Research Council,
<irc.nrc-cnrc.gc.ca/ie/lighting/daylight/dds.htlm>; March 2010.
[4] DIALUX 4.0. Germany: DIAL GmbH; 2005.
[5] Bellia L, Cesarano A, Minichiello F, Sibilio S. De-Light: a software tool for the eval-
uation of direct daylighting illuminances both indoors and outdoors e comparison
with Superlite 2.0 and Lumen Micro 7.1. Building and Environment
2000;35:281e95.
[6] De Rosa A, Ferraro V, Igawa N, Kaliakatsos D, Marinelli VINLUX. A calculation
code for daylight illuminance predictions inside buildings and its experi-
mental validation. Building and Environment 2009;44:1769e75.
[7] Ward G. The RADIANCE lighting simulation and rendering system. Proceed-
ings of ’94 SIGGRAPH conference. Computer Graphics; 1994:459e72.
[8] Ward G, Shakespeare R. Rendering with RADIANCE. The art and science of
lighting visualization. California: Space and Light; 2003.
[9] SUPERLITE 2.0. Predicting daylighting and lighting performance in complex
building spaces. USA: University of California, Lawrence Berkeley Laboratory;
1994.
[10] Robinson D, Stone A. Internal illumination prediction based on a simplified
radiosity algorithm. Solar Energy 2006;80:260e7.
[11] Juslén H, Tenner A. Mechanisms involved in enhancing human performance
by changing the lighting in the industrial workplace. International Journal of
Industrial Ergonomics 2005;35:843e55.
[12] Galasiu AD, Veitch JA. Occupant preferences and satisfaction with the lumi-
nous environment and control systems in daylit offices: a literature review.
Energy and Buildings 2006;38:728e42.
[13] Manav B. An experimental study on the appraisal of the visual environment at
offices in relation to colour temperature and illuminance. Building and
Environment 2007;42:979e83.
[14] Durak A, Olguntürk NC, Yener C, Güvenç D, Gürçınar Y. Impact of lighting
arrangements and illuminances on different impressions of a room. Building
and Environment 2007;42:3476e82.
[15] França F, Howell J, Ezekoye O, Morales JC. Inverse design of thermal systems.
Advances in Heat Transfer 2002;36:1e110.
[16] Ertürk H, Ezekoye OA, Howell JR. Comparison of three regularized solution
techniques in a three-dimensional inverse radiation problem. Journal of
Quantitative Spectroscopy and Radiative Transfer 2002;73:307e16.
[17] França FHF, Howell JR. Transient inverse design of radiative enclosures for
thermal processing of materials. Inverse Problems in Science and Engineering
2006;14:423e36.
[18] Daun K, França FHR, Larsen M, Leduc G, Howell JR. Comparison of methods for
inverse design of radiant enclosures. Journal of Heat Transfer 2006;128:
269e82.
[19] Smith Schneider P, Mossi AC, França FHR, De Sousa FL, Silva Neto AJ. Appli-
cation of inverse analysis to illumination design. Inverse Problems in Science
and Engineering 2009;17:737e53.
[20] De Sousa FL, Ramos FM, Paglione P, Girardi RM. New stochastic algorithm for
design optimization. AAIA Journal 2003;41:1808e18.
[21] Hansen C. Rank-deficient and discrete ill-posed problems: numerical aspects
of linear inversion. Philadelphia: SIAM; 1998.
[22] Silva Neto AJ, Moura Neto FD. Inverse problems e fundamental concepts and
applications. EdUERJ; 2005 (in Portuguese).
[23] Knupp DC, Silva Neto AJ, Sacco WF. Radiative properties estimation with the
particle collision algorithm based on a sensitivity analysis. High Tempera-
tures e High Pressures 2009;38:137e51.
[24] Knupp DC, Silva Neto AJ, Sacco WF. Radiative properties estimation with the
Luus-Jaakola and the particle collision algorithm. Computer Modeling in
Engineering & Sciences 2009;54:121e46.
[25] Cuco APC, Silva Neto AJ, Campos Velho HF, Sousa FL. Solution of an inverse
adsorption problem with an epidemic genetic algorithm and the generalized
extremal optimization algorithm. Inverse Problems in Science and Engi-
neering 2009;17:289e302.
[26] Stephany S, Becceneri JC, Souto RP, Campos Velho HF, Silva Neto AJ. A pre-
regularization scheme for the reconstruction of a spatial dependent scattering
albedo using a hybrid ant colony optimization implementation. Applied
Mathematical Modelling 2010;34:561e72.
[27] Lobato FS, Steffen Jr V, Silva Neto AJ. Solution of inverse radiative transfer
problems in two-layer participating media with differential evolution. Inverse
Problems in Science and Engineering 2010;18:183e95.
[28] Sousa FL, Soeiro FJCP, Silva Neto AJ, Ramos FM. Application of the generalized
extremal optimization algorithm to an inverse radiative transfer problem.
Inverse Problems in Science and Engineering 2007;15:699e714.
[29] Galski RL, Sousa FL, Ramos FM, Silva Neto AJ. Application of a GEOþSA
hybrid optimization algorithm to the solution of an inverse radiative
transfer problem. Inverse Problems in Science and Engineering 2009;17:
321e34.
[30] Boettcher S, Percus AG. Optimization with extremal dynamics. Physical
Review Letters 2001;86:5211e4.