Automation in Construction 11 Ž2002.

173–184
www.elsevier.comrlocaterautcon

A design optimization tool based on a genetic algorithm

Luisa Gama Caldas ) , Leslie K. Norford
Massachusetts Institute of Technology, Room 5-418, 77 Massachusetts AÕenue, Cambridge, MA 02139, USA

Abstract

Much interest has been recently devoted to generative processes in design. Advances in computational tools for design
applications, coupled with techniques from the field of artificial intelligence, have lead to new possibilities in the way
computers can inform and actively interact with the design process.
In this paper, we use the concepts of generative and goal-oriented design to propose a computer tool that can help the
designer to generate and evaluate certain aspects of a solution towards an optimized behavior of the final configuration. This
work focuses mostly on those aspects related to the environmental performance of buildings.
Genetic Algorithms ŽGAs. are applied as a generative and search procedure to look for optimized design solutions in
terms of thermal and lighting performance in a building. The GA is first used to generate possible design solutions, which
are then evaluated in terms of lighting and thermal behavior using a detailed thermal analysis program ŽDOE2.1E.. The
results from the simulations are subsequently used to further guide the GA search towards finding low-energy solutions to
the problem under study. Solutions can be visualized using an AutoLisp routine.
The specific problem addressed in this study is the placing and sizing of windows in an office building. The same method
is applicable to a wide range of design problems like the choice of construction materials, design of shading elements, or
sizing of lighting and mechanical systems for buildings. q 2002 Elsevier Science B.V. All rights reserved.

Keywords: Artificial intelligence; Genetic algorithms; Building design; Low–energy design; Generative systems; Optimization in architec-
ture; Architectural design; Genetic algorithms in architecture; Artificial intelligence in architecture; Environmental design

1. Introduction it possible to fully understand all the factors involved

in the process. By using simulation tools, it is possi-
Computer simulations have proven in the last few ble to engage in a design practice based on feedback
decades to be a powerful tool for studying the envi- loops between making design decisions and evaluat-
ronmental performance of buildings. The interaction ing their environmental impact, as a way to inform
between design features, climate, occupants and me- the on-going process of design.
chanical and electrical systems in a building is a The objective of this paper is to go a step further
complex one, and only by resorting to simulations is in integrating the computer media in the design
process by making use of it not only as a simulation
tool, but as a generative method too. Simulations are
normally used in a scenario-by-scenario base, with
)
Corresponding author. Tel.: q351-919-907-852; fax: q351-
the designer generating a solution and subsequently
213-861-686. having the computer evaluating it. This is however a
E-mail address: [email protected] ŽL.G. Caldas.. slow and tedious process and typically only a few

0926-5805r02r$ - see front matter q 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 6 - 5 8 0 5 Ž 0 0 . 0 0 0 9 6 - 0
174 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184

scenarios are evaluated from within a large range of design, like lighting levels within the space, internal
possible choices. In opposition to this method, we temperatures, and amount of energy spent in light-
propose an approach of goal-oriented design w13x, ing, heating and cooling.
where the computer is used to extensively search the In this study, we propose an approach where the
design solution space looking for high performance evaluation criteria are related to the environmental
solutions in terms of specified goals. This way the performance of the building, both in terms of light-
computer automatically generates and evaluates pos- ing and thermal behavior. The tool described is
sible configurations, and presents the designer with mostly aimed at intermediate to late stages of design,
optimal or near-optimal solutions for the problem since at early stages of design there is often no
under study. optimum solution, instead there is a large range of
The use of optimization rather than simulation possible solutions all potentially having a good per-
w17x is thus proposed as a way to integrate the formance in responding to the problem under consid-
computer media in the design process. The method eration. As the design process progresses, there is
used to implement the idea of goal-oriented design however an increased opportunity for optimization,
was the application of a search and optimization for the solution becomes increasingly more defined
technique borrowed from the field of artificial intelli- and more concrete requirements are asked of its
gence, Genetic Algorithms ŽGAs.. A GA is a proce- components. Examples of addressable problems are
dure loosely based on Darwinian notions of survival sizing of glazing elements according to climate and
of the fittest, where selection and recombination orientation, choice of type of glazing, choice of
operators are used among candidate solutions to look construction materials for walls and roofs, type and
for high performance ones w6x. We use GAs to search sizing of shading elements, and many others.
for design solutions, evaluate them and continue the
search guided by the results of that evaluation. The
designer is then provided with the results of that
looping process. 3. Problem under study
Two important steps in the design process are the
generation of a solution to the problem under study, The particular problem we present in this paper is
and the evaluation of how good a response it repre- the optimal sizing of windows in a building to
sents to that problem. The field of artificial intelli- optimize its lighting, heating and cooling perfor-
gence provides the designer with novel ways to mance. The exact sizing of windows occurs typically
generate and evaluate design solutions. By allowing at later stages of design, but will have a significant
the computational media to actively interact with and effect on the environmental behavior of a building.
inform the design process, it is expected that solu- The optimal sizing will depend on the climate the
tions that were not immediately obvious to the de- building is located in, the glazing used, the orienta-
signer may be generated. Throughout this process, tion the window is facing, and the type of use of the
the design object is enriched and expanded by its building Žoffices, housing, etc...
interaction with the media in which it develops. Windows display a complex environmental be-
havior due to different requirements often being in
conflict with each other, making the evaluation of
their overall performance a complex one. During the
2. GA-based design tool heating season, as window sizes increase, there is an
increase in daylighting availability and thus a de-
Part of the difficulty of using generative tools to crease in the need for artificial lighting, leading to
optimize architectural design is defining proper eval- less electrical energy consumption. There are also
uation criteria for a given solution. Environmental more useful solar gains through the window that
analysis is a particularly adequate field for goal-ori- tends to reduce heating expenditure. On the other
ented design because it is possible to specify quan- hand, there is an increase in heat loss through the
tifiable performance criteria to be achieved with the glass, which in turn requires more spending in heat-
 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184 175

ing. Less use of artificial lighting will also lead to a layer of spectrally selective glass, with a shading
lower internal heat gains in the building, leading to coefficient of 0.34 and visual transmittance of 0.41.
an increased need of heating. The problem was studied for two locations:
During the cooling season, as window sizes in- Phoenix, Arizona, a cooling-dominated situation; and
crease, there is a decrease in the need for artificial Chicago, Illinois, a heating-dominated climate. The
lighting, which leads to less electrical energy con- aim is to provide some insight on how the optimal
sumption and also to lower cooling loads generated sizing of windows varies with different climatic con-
by the lights themselves. On the other hand, there are ditions.
increased solar gains leading to more spending in
cooling.
These interactions are simulated in the following 4. GAs, description and application
way: given a design configuration, for each hour in
the year daylighting levels inside the space are calcu- We now further describe the optimization algo-
lated at reference points. These numbers are then rithm used. A GA is a global search technique
compared to the required lighting levels set by the adequate for searching noisy solution spaces with
designer. The artificial lighting system, which is local and global minima. Because it searches from a
continuously dimmable, is set to provide just enough population of points, not a single point, the probabil-
light to make up from the difference between avail- ity of the search getting trapped in a local minimum
able daylighting and required light levels. The elec- is limited. GAs start searching by randomly sampling
trical energy spent by the lights is quantified, and so within the solution space and then use stochastic
are the heat gains released into the space by them. operators to direct a hill-climbing process based on
The system then goes on to calculating internal objective function values w6x.
temperatures inside the space by taking into account Under GA terminology, a solution to a problem is
external climatic conditions, the characteristics of the an indiÕidual and the group of solutions existent at
building fabric Žwall, roof and window materials., each stage is a population. Each time a new popula-
solar gains through the windows, internal gains from tion of individuals is created, it is called a genera-
occupants, equipment and artificial lighting, etc. tion. In binary GAs like the one used in this study,
Those temperatures are compared to set points de- each individual is represented by a binary string
fined by the designers, and heating or cooling is called a chromosome, which codes all the parameters
provided if required to compensate for the difference of interest corresponding to that individual. A chro-
between calculated temperatures and predefined set- mosome is formed of alleles, the binary coding bits.
points. The energy spent in heating and cooling is The fitness of any particular individual corresponds
determined, and added to lighting energy to deter- to the value of the objective function at that point.
mine total annual energy consumption in the build- Genetic operators control the evolution of the
ing. generations of problem solutions. The three basic
The building simulated is assumed to be an office genetic operators are reproduction, crossover and
building, where internal conditions are usually tightly mutation. The probability of a given solution being
controlled, and systems like dimmable artificial light- chosen for reproduction is proportional to the fitness
ing are likely to exist. The hypothetical office build- of that solution. Crossover implies that parts of two
ing module used consists of a square core zone, 30.5 randomly chosen chromosomes will be swapped to
m on a side, surrounded by four identical perimeter create a new individual. Mutation involves randomly
zones, each of 30.5 = 4.6 m. The module faces the changing an allele in a solution to look for new
four cardinal directions. Each perimeter zone is di- points in the solution space. Although there are more
vided into 10 office spaces of equal size Ž3.1 = 4.6 elaborate versions of these operators, the basic prin-
m.. The reference point for daylighting calculation is ciples remain similar for most GAs.
located 3.1 m from the window, at desk height. The A GA starts by generating a number of possible
required horizontal illuminance at this point is 540 solutions to a problem, evaluates them and applies
lx. The window glazing was set as double pane with the basic genetic operators to that initial population
176 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184

according to the individual fitness of each individual. tal level, with selection and cross-over being per-
This process generates a new population with higher formed manually by the designer w14x.
average fitness than the previous one, which will in Some other optimization methods have also been
turn be evaluated. The cycle will be repeated for the proposed for applications to building related prob-
number of generations set by the user, which is lems. Monks et al. w13x used a combination of simu-
dependent on problem complexity. Despite their ap- lated annealing and steepest descent to optimize
parent simplicity, GAs have proved to have high parameters related to surface geometry and choice of
efficacy in solving complex problems for which materials to achieve a set of acoustical based goals.
conventional hill-climbing derivative-based algo- Radford and Gero w16,17x proposed the use of dy-
rithms are likely to be trapped in local solutions. namic programming to optimize window sizing and
Typical population sizes for GAs range from 30 glazing materials to reduce building energy con-
to 200, based on earlier studies such as those of sumption.
Grefenstette w7x, where suggestions for optimal popu-
lation choices based on parametric studies are pre-
sented. In this study, we use a strategy named micro- 5. Thermal simulation method
GA w11x, which starts with a small population Žin this
case, of only five individuals. and quickly makes it The thermal and lighting simulations are per-
converge to a solution. Convergence is measured by formed using DOE2.1E, a popular hourly thermal
comparing the chromosomes of the individual solu- simulation program w19x. It calculates daylighting
tions. If they differ by less that 5%, it is considered based on the daylight factor method w18,23x. Previ-
that the population has converged. When that hap- ous attempts to use DOE-2.1E to optimize design
pens, the micro-GA generates a new random popula- parameters of buildings include using regression
tion while carrying over the individual with the best analysis on data created by parametric DOE-2 runs
fitness in the previous generation — a strategy known w21x and the use of indices and weighting factors
as elitism. This way, new individuals are often w20x.
brought into the search, without loosing track of the The method used in this study was to create a link
ones that did better until that point. An advantage of between the GA and DOE2.1-E, on the UNIX plat-
using the micro-GA procedure is that the algorithm form, and make a call to DOE2.1-E each time it is
tends to perform a local search around the best necessary to calculate the fitness of an individual.
solutions during the generations prior to conver- The objective function of a solution is thus the result
gence, since at that stage solutions only differ by a of running a DOE2.1-E thermal simulation of the
few alleles. This local search is important in finding building under study, with the parameters determined
local minima around good solutions, and is usually by the individual chromosome of that solution.
hard to implement in conventional GAs. AutoLisp procedures allow the results to be visu-
GAs have been used in building applications re- ally inspected as AutoCad drawings, both two- and
lated to energy consumption, mostly to optimize the three-dimensionally. The designer can in this way
sizing and control of HVAC systems w3,8,24x. At the visualize the several steps of evolution of the pro-
structural analysis level, GAs have been used for cess, as well as the final results.
design optimization of trusses w1,4,6x, beams w12,22x
and columns w9x. Several applications of evolutive
design to spatial problems have been found for floor 6. Description of hand method
layout problems, where GAs are used to assign uses
to specific areas in a building w2,5,10,15x. Attempts To test if the proposed method would actually be
have also been made to apply GAs to the design of able to locate optimal solutions for the problem
three-dimensional shapes, to help designers at early under study, the algorithm was first tested against a
stages of design. The difficulty of finding appropri- hand-worked example of limited size, for which the
ate objective functions for design evaluation has optimal solution was found by sampling all possible
made some of these attempts to remain at experimen- solutions.
 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184 177

Fig. 1. Annual energy consumption ŽMWh. for south window Fig. 3. Annual energy consumption ŽMWh. for west window
dimensions. dimensions.

The location chosen for the testing was Phoenix, 0.9 m high. For each orientation, a 100-generation
AZ. We first determined the solution space for each GA was run.
of the four cardinal directions. We considered eight The solution spaces for each of the four orienta-
window widths and heights for each orientation, tions are shown in Figs. 1–4. In general, the solution
from 0.3 to 2.4 m at discrete steps of 0.3 m, creating spaces proved to have several local minima and a
a solution space of 64 points. While the dimensions global minimum for all orientations but north. There
for an orientation were varied, the windows facing are no local minima for the north orientation. The
the other orientations were kept 1.2 m wide and presence of local minima makes the problem inap-
propriate for derivative-based search methods and
makes GA a reasonable approach.

Fig. 2. Annual energy consumption ŽMWh. for east window Fig. 4. Annual energy consumption ŽMWh. for north window
dimensions. dimensions.
178 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184

7. Results and discussion local minima correspond to windows with the same
area Ž1.9 m2 ., but with different aspect ratios that
Fig. 1 shows the solution space for south-facing show the effect of accounting for daylighting pene-
windows for the Phoenix climate. It can be seen that tration in the space. The more horizontal window
there is a relatively flat surface of configurations corresponds to a slightly lower energy consumption.
corresponding to low energy consumption. Within The optimal area is 5% larger than the area of the
that flat surface, there are however several local second local minimum.
minima and a global minimum. Being stuck in a For the east orientation, the global minimum was
local minimum would not be too serious in this case similar to those for the south orientation Ž2.1, 0.9.,
since the objective function difference in relation to and the local minima were located in points Ž1.8,
the global minimum is small. However, we are inter- 0.9. and Ž1.2, 1.5.. The GA located point Ž1.8, 0.9.
ested in the global algorithm behavior and not only by generation 16, but did not manage to locate point
in the particular results for this example. Ž2.1, 0.9. in 100 generations. Point Ž1.8, 0.9. stands
The global minimum of 116.94 MWh is located at next to the global minimum, and has an objective
point Ž2.1, 0.9., corresponding to a window 2.1 m function value only 0.03 MWh higher than it. The
wide and 0.9 m high. There were two local minima observed behavior is a consequence of the fact that
of 117 MWh for points Ž1.2, 1.5. and Ž0.9, 2.1.. The GAs have a difficulty in succeeding in very local
GA located the global minimum in generation 18, searches. GAs are global search procedures and are
and the Ž0.9, 2.1. local minimum in generation 21. It often able to locate the global minimum for the
can be seen that the global minimum and one of the problem. However, the search can get very close to

Fig. 5. Initial solutions for Phoenix climate, experiment 3. Each row represents an individual solution. The numbers on the right show
annual energy consumption in MWh for each solution.
 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184 179

the global minimum without actually finding it. It 8. Extension of the method to a larger solution
should be noted that the actual difference in objec- space
tive function values is negligible in this case.
For the west orientation, the solution landscape The resort to a hand-worked example proved the
proved to be flatter than for south and east orienta- GA succeeded in most cases in finding a global
tions. Although solar angles are symmetrical for the minimum for the problem. In the cases it did not
east and west orientations, temperatures and occu- reach the actual minimum it localized points very
pancies are not. The global minimum was at point closed to it, which provides enough confidence in the
Ž1.5, 0.9. and local minima were found at points results to extend the method to larger problems for
Ž0.6, 2.1., Ž0.9, 1.5. and Ž2.4, 0.6.. The GA located which it is not possible to calculate manually the
the global minimum by generation 10. solution space, given its size.
For the north orientation, there was only a global The GA was then used to search in a hyperspace
minimum at point Ž2.1, 2.1., with no other local of eight dimensions for the optimal size of windows
minima. The GA located the global minimum very to all orientations simultaneously, resulting in a solu-
quickly, at generation 7. This good performance of tion space of 16,777,216 points. For the experimen-
the GA is thought to be due to the smooth configura- tal building under study, it is possible to know the
tion of the solution space in this case. Apart from the global minimum for the eight-dimensional space by
minimum, the GA located a series of high-perfor- combining the individual minima for each orienta-
mance solutions in its vicinity, all by the end of tion. This is due to the way the building was set up,
generation 8. with no interactions between the different orienta-

Fig. 6. Final solutions for Phoenix climate, experiment 3. Each row represents an individual solution. The numbers on the right show annual
energy consumption in MWh for each solution.
180 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184

tions either in terms of space or building systems. with an annual electrical energy consumption of
The trial building was designed in this way so that 131.4 MWh, about 15% higher than the best possible
the objective function would be separable, allowing result.
the global minimum to be known to better test the Figs. 5 and 6 show the evolution for one experi-
performance of the algorithm. However, regular ment from the initial solutions to the final generation
buildings are not always designed this way and thus of optimized designs. It can be seen that initially
it is usually not possible to optimize for each direc- there is a high variance in the design patterns that
tion separately and then combine the individual re- progressively converges into the optimized solutions.
sults. Fig. 7 shows the evolution of design generations
From the hand-worked example, it is possible to to the Phoenix climate for another experiment. For
know that for Phoenix climate, the best possible reasons of space, only every tenth generation is
result corresponds to the point Ž2.1, 2.1, 2.1, 0.9, 2.1, shown here, until generation 80 when the solution
0.9, 1.5, 0.9. Žfirst listing north window dimensions has converged. It can be seen that although many
— width, height — then south, east and west., different configurations are tested along the search,
corresponding to an annual electrical energy con- the design steadily tends to the final optimized solu-
sumption of 114.53 MWh. The worst point in the tion: large windows for north orientation Ž2.1 = 2.1
hyperspace is Ž0.3, 0.3, 2.4, 2.4, 2.4, 2.4, 2.4, 2.4., m., where available daylighting can be maximized

Fig. 7. Generations 1–80 for Phoenix climate, experiment 1.

 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184 181

Table 1
Solutions for Phoenix climate
Experiment Best solution found Objective function Percentage above
value ŽMWh. minimum Ž%.
1 2.1r2.1r1.2r1.5r1.2r1.5r1.2r0.9 114.74 0.018
2 2.4r2.1r2.1r0.9r2.1r0.9r0.9r1.5 114.65 0.01
3 2.1r2.1r1.2r1.5r2.1r0.9r1.5r0.9 114.60 0.006
4 1.8r2.1r1.2r1.8r0.9r1.5r1.2r0.9 114.89 0.03
5 2.4r2.1r2.1r0.9r1.2r1.5r1.5r0.9 114.65 0.01

without incurring on high solar heat gains in the sive west gains, will lead to internal overheating.
summer. Due to the mild winter climate and the use These solutions are dependent on the type of glazing
of double pane windows, heat losses through this used: a different glazing option would probably lead
large glazing area are not enough to penalize this to different optimum window sizes.
solution when the overall energy consumption is In total, five experimental runs were made for
calculated. Average size windows Ž1.2 = 1.5 m. are Phoenix climate, to compare results from different
chosen for east and south orientations, showing a trials. A maximum number of 100 generations were
balance between admitting useful daylighting and used for each experiment. The results are summa-
winter solar gains and preventing too much solar rized in Table 1. In general, the GA performed well,
gain in the summer. Smaller windows Ž1.2 = 0.9 m. being able to find high-performance solutions close
are chosen for west since that is the worse orienta- to the global minimum. The results are even more
tion in terms of summer overheating: the building is satisfactory considering that 100 generations is a
already warm at the end of the afternoon and exces- relatively low number for GAs, and the population

Fig. 8. Comparative results for Phoenix and Chicago locations.

182 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184

Table 2
Solutions for Chicago climate
Experiment Best solution found Objective function
value ŽMWh.
1 0.3r0.3r1.8r1.5r1.2r1.5r2.4r0.3 133.03
2 0.3r0.3r1.5r1.5r0.9r1.5r1.5r0.9 132.97
3 0.6r0.3r1.5r1.5r1.2r0.9r0.3r0.9 133.15
4 0.3r0.3r1.8r1.5r1.8r0.9r0.9r0.6 133.00
5 0.3r0.3r1.5r2.4r1.8r1.2r0.6r0.3 133.03

size of 5 is very small. The maximum of 500 differ- the same climate, suggest the use of an optimization
ent solutions evaluated by the GA corresponds to tool like the one proposed here may not only provide
only 0.00003% of the solution space dimension. increased energy savings but also introduce a posi-
Results from these experiments are also displayed tive degree of variability in the design. It also sug-
in Fig. 8. It can be seen that although total energy gests that unexpected solutions that might not be
consumption is very similar, design solutions are obvious to the designer may be pointed out by the
somewhat different. This leads to the conclusion that computer, such as the case of the very large north-
different runs of the optimization tool may lead to facing windows for Phoenix, or the horizontal strip
different solutions with similar performance, a fact facing west for Chicago.
that may be of interest to the designer, who may try Solutions optimized for environmental perfor-
more than one run of the GA to get a sense of mance may not always represent optimal behavior in
available alternatives. terms of other criteria. For example, the north-facing
Five similar experiments were also done for windows for the Chicago climate may be found too
Chicago climate, a cooling dominated situation, to small to provide occupants with an adequate view. In
assess to what extent optimum window designs may those cases, the designer may modify the results
be sensitive to climatic conditions. In this case, the provided by the GA while making use of the knowl-
global minimum was not known in advance and so it edge gained during the optimization process, such as
was not possible to determine the margin by which knowing that north-facing windows should be kept
the solutions found differed from the optimum solu- as small as possible. A simple computer simulation
tion. Results are summarized in Table 2. can then be performed to determine to what extent
Fig. 8 also shows the solutions obtained for the modified design’s energy consumption differs
Chicago climate, which vary in significant ways from the optimized solution.
from Phoenix solutions. Analyzing the results of the
first experiment shows us that due to the extreme
9. Conclusions
winter conditions, the north window size was re-
duced to the minimum allowed Ž0.3 = 0.3 m., since This paper presents a generative tool to optimize
north windows are in this case a large source of heat design elements of a building in terms of their
losses and provide little solar gains in winter. The environmental performance. It makes use of a GA as
south window size increased in relation to the the search engine, a thermal and lighting simulation
Phoenix solutions, to allow for more solar gains in program, and an AutoLisp routine for visualization
winter Ž1.8 = 1.5 m.. East window size remained of results.
basically similar Ž1.2 = 1.5 m., and west was re- The method proposed was first validated in rela-
duced to a horizontal strip Ž2.4 = 0.3 m. that allows tion to a hand-worked example for which the optimal
for some daylighting but prevents excessive heat solution could be calculated manually. Results proved
loss. highly satisfactory and provided enough confidence
The large differences between the solutions found for the process to be extended to a larger solution
for each climate, and also between orientations within space for which there is no practical way of calculat-
 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184 183

ing the optimized solutions. The extension of the References

method was applied to a hyperspace of eight dimen-
sions, corresponding to simultaneously sizing eight w1x C. Camp, S. Pezeshk, G. Cao, Optimized design of two-di-
window dimensions in the building. mensional structures using a genetic algorithm, Journal of
The results obtained showed that from initial pop- Structural Engineering 124 Ž5. Ž1998. 551–559.
ulations with high diversity, the algorithm slowly w2x J. Damsky, J. Gero, An evolutionary approach to generating
converged to populations consisting of optimal or constraint-based space layout topologies, in: R. Junge ŽEd..,
CAAD Futures 1997, Kluwer Academic Publishing, Dor-
near-optimal solutions. For different runs of the drecht, 1997, pp. 855–874.
problem, the solution obtained can have some varia- w3x S. Dickinson, A. Bradshaw, Genetic algorithm optimization
tion, which shows that different configurations may and scheduling for building heating systems, Genetic Algo-
correspond to similar standards of environmental rithms in Engineering Systems: Innovations and Applica-
performance. This may constitute valuable informa- tions, 12–14 September, 1995, pp. 106–111, University of
Sheffield: Conference Publication No. 414, Institution of
tion to the designer, who is provided with a number Electrical Engineers.
of alternative solutions over which he can further w4x M. Galante, Genetic algorithm as an approach to optimize
overlap other design criteria not included in the real-world trusses, International Journal for Numerical Meth-
optimization process. It was also observed that for ods in Engineering 39 Ž1996. 361–382.
w5x J. Gero, V. Kazakov, Evolving design genes in space layout
different orientations the solutions achieved varied
problems, Artificial Intelligence in Engineering 12 Ž3. Ž1998.
substantially, as expected from the different require- 163–176.
ments the glazing elements had to respond to. w6x D. Goldberg, Genetic Algorithms in Search. Optimization
It was shown that for different climates the design and Machine Learning, Addison-Wesley Publishing, 1989.
solutions proposed by the algorithm differed consid- w7x J. Grefenstette, Optimization of control parameters for ge-
erably. The same building layout needs to have netic algorithms, IEEE Transactions on Systems, Man and
Cybernetics SMC-16 Ž1. Ž1986. 122–128.
different design components according to the climate w8x W. Huang, H. Lam, Using genetic algorithms to optimize
it is located in order to optimize its environmental controller parameters for HVAC systems, Energy and Build-
performance. This optimization cannot be achieved ings 26 Ž1997. 277–282.
in a scenario-by-scenario basis due to the size of the w9x R. Ishida, Y. Sugiyama, Proposal of constructive algorithm
problems involved, but it is possible to perform and discrete shape design of the strongest column, AIAA
Journal 33 Ž3. Ž1995. 401–406.
using a search procedure like the one proposed in w10x J. Jo, J. Gero, Space layout planning using an evolutionary
this study. approach, Artificial Intelligence in Engineering 12 Ž3. Ž1998.
The problem presented here is just one example 149–162.
of the way goal-oriented design for environmental w11x K. Krishnakumar, Micro-genetic algorithms for stationary
performance can be applied to inform the design and non-stationary function optimization, in: G. Rodriguez
ŽEd.., Intelligent Control and Adaptive Systems, 7–8
process. The same procedure of generation and eval- November, SPIE — The International Society for Optical
uation of different solutions guided by a search Engineering, Philadelphia, PA, 1989, pp. 289–296.
algorithm can also be applied to a variety of other w12x J. Marcelin, P. Trompette, R. Dornberger, Optimization of
problems in building design. composite beam structures using a genetic algorithm, Struc-
tural Optimization 9 Ž3–4. Ž1995. 236–244.
w13x M. Monks, B. Oh, J. Dorsey, Audioptimization: Goal based
acoustic design, MIT Technical Report MIT-LCS-TM-588,
Acknowledgements
1998.
w14x U. O’Reilly, G. Ramachandran, A preliminary investigation
This study was developed as part of a Ph.D. dis- of evolution as a form design strategy, in: C. Adami, R.
sertation in the Building Technology Program, De- Belew, H. Kitano, C. Taylor ŽEds.., Proceedings of the Sixth
partment of Architecture, Massachusetts Institute of International Conference on Artificial Life, MIT Press, 1998.
Technology, with funding from Fundaçao ˜ para a w15x D. Pham, H. Onder, A knowledge-based system for optimiz-
ing workplace layouts using a genetic algorithm, Ergonomics
ˆ
Ciencia e Tecnologia, Programa PRAXIS XXI, Por-
35 Ž12. Ž1992. 1479–1487.
tugal. We would like to thank Stephen Selkowitz, w16x A. Radford, J. Gero, On the design of windows, Environment
Robert Sullivan and Eleanor Lee at Lawrence Berke- and Planning B 6 Ž1. Ž1978. 41–45.
ley National Laboratory. w17x A. Radford, J. Gero, A dynamic programming approach to
184 L.G. Caldas, L.K. Norfordr Automation in Construction 11 (2002) 173–184

the optimum lighting problem, Engineering Optimization 3 solar control and daylighting performance in commercial
Ž2. Ž1978. 71–82. office buildings, ASHRAErDOErBTECC Conference on
w18x S. Selkowitz, J. Kim, M. Navvab, F. Winkelman, The DOE-2 the Thermal Performance of the Exterior Envelopes of Build-
and Superlite daylighting programs, Passive 82, The Natio- ings V, December 7–10, Clearwater Beach, FL, 1992.
nal Passive Solar Conference, August 29–September 3, w22x B. Wang, J. Chen, Applications of genetic algorithm for the
Knoxville, TN, 1982. support location optimization of beams, Computers and
w19x Simulation Research Group, DOE-2 Supplement — Version Structures 58 Ž4. Ž1996. 797–800.
2.1E, Lawrence Berkeley National Laboratory, LBL-34946, w23x F. Winkelman, 1993, Daylighting Calculation in DOE-2,
1993. Lawrence Berkeley National Laboratory Report LBL-11353,
w20x R. Sullivan, D. Arasteh, K. Papamichael, J. Kim, R. Johnson, May 1983.
S. Selkowitz, R. McCluney, An indices approach for evaluat- w24x J. Wright, HVAC optimization studies: Sizing by genetic
ing the performance of fenestration systems in nonresidential algorithm, Building Services Engineering Research and
buildings, ASHRAE Transactions 94 ŽPart 2. Ž1988.. Technology 17 Ž1. Ž1996. 1–14.
w21x R. Sullivan, E. Lee, S. Selkowitz, A method for optimizing