See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/315176145

Multiobjective Design Optimization of Building Space Layout, Energy, and

Daylighting Performance

Article  in  Journal of Computing in Civil Engineering · March 2017

DOI: 10.1061/(ASCE)CP.1943-5487.0000669

CITATIONS READS
7 613

2 authors:

Ipek Gursel Dino Göktürk Üçoluk

Middle East Technical University Middle East Technical University
44 PUBLICATIONS   176 CITATIONS    40 PUBLICATIONS   599 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Mixed Reality Applications and Serious Games in Architectural Education View project

Design Computing Summer School 2018 View project

All content following this page was uploaded by Ipek Gursel Dino on 23 October 2017.

The user has requested enhancement of the downloaded file.

 Multiobjective Design Optimization of Building Space
Layout, Energy, and Daylighting Performance
Ipek Gürsel Dino 1 and Göktürk Üçoluk 2

Abstract: Computational tools for early architectural design need to address issues related to building performance and integrally con-
sider early design decisions regarding building form, spatial layout, orientation, and envelope articulation. This paper presents (1) a design
optimization tool, Multiobjective Architectural Design Explorer (MADE) that supports performative-based building design, and (2) a
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

design exploration strategy that effectively operationalizes MADE for Pareto-based comparative search. MADE implements genetic
optimization in two steps. It first generates building layouts that satisfy formal, topological, and placement constraints with a single-
objective genetic algorithm. Then, MADE determines the opening sizes of the generated layout(s) by optimizing the buildings’ energy
and daylighting performance with a multiobjective genetic algorithm. MADE is tested on the design of a library building and the proposed
Pareto-based exploration strategy is demonstrated in the selection of an optimal design. The results point out the feasibility of the presented
tool and the design strategy, and that multiple objectives can be satisfied by sequentially optimizing design objectives. DOI: 10.1061/
(ASCE)CP.1943-5487.0000669. © 2017 American Society of Civil Engineers.
Author keywords: Building energy and daylighting optimization; Architectural layout optimization; Genetic algorithms; Design
exploration; Pareto front.

Introduction with design tools for quantifying performance criteria. In this

regard, the conventional role of performance simulation as post-
Early architectural design plays a key role in ensuring building per- rationalization can be reversed, such that performance analysis is
formance. The decisions taken during this phase determine up to deeply rooted into design synthesis to facilitate well-informed
80% of a building’s environmental load and operation costs decisions being taken. Specifically, genetic optimization inte-
(Bogenstätter 2000), so there is a growing demand for tools grated with simulation tools can automate the generation and
and methods supporting the design of high-performance build- evaluation cycle and facilitate the analysis and comparison of
ings. Early performance-based design decisions relate to various many design alternatives.
criteria regarding building form and orientation as well as the In this paper, a design tool named Multiobjective Architectural
envelope openings (Bleil de Souza and Tucker 2015). Addition- Design Explorer (MADE) and a design strategy for operationaliz-
ally, its spatial layout has an influence on a building’s thermal and ing MADE are presented. MADE supports performative building
daylighting characteristics due to the spaces’ heat gain and solar design by implementing genetic optimization in two steps in two
exposure. Therefore, tools targeting early design should integrally corresponding modules, MADE-Layout and MADE-Opening.
consider a building’s space layout design as part of energy-related MADE-Layout first optimizes the design of building layouts with
factors. a novel heuristics developed by the authors and a single-objective
Computation and quantitative analytical methods are impor- genetic algorithm. A detailed discussion on the development and
tant in supporting performance-based decision making, especially evaluation of this module was presented in Dino (2016). The sec-
in cases of low-information granularity and a multitude of objec- ond module, MADE-Opening, determines the opening sizes of the
tives that conflict with each other. Metaheuristic optimization has generated layouts by the multiobjective optimization of building
the potential to provide the user with a deeper insight into the energy and daylighting performance using a performance simula-
design space while considering the performance of alternative tion tool, EnergyPlus. MADE is eventually tested though a case
solutions with respect to a number of objectives. Specifically, study that aims to demonstrate how it can effectively support per-
genetic algorithms can support multivariate problem solving with formative design exploration. For this, design studies are carried
multimodal and discontinuous design spaces, and address creative out that comparatively evaluate different building forms, layouts,
design by maintaining a balanced amount of divergence and con- and orientation for building instances or Pareto fronts. Each are
vergence. Additionally, building simulation tools can be coupled quantified by means of their objective functions and hypervolume
(HV), facilitating benchmarking toward better-informed decision
1
Assistant Professor, Dept. of Architecture, Middle East Technical making. The proposed exploration strategy is aimed to set an
Univ., Universiteler Mahallesi, Dumlupinar Bulvari, No. 1, Ankara example of design processes that systematically explore the solu-
06800, Turkey (corresponding author). E-mail: [email protected] tion space with MADE for better-informed decision making.
2
Professor, Dept. of Computer Engineering, Middle East Technical
Univ., Universiteler Mahallesi, Dumlupinar Bulvari, No. 1, Ankara
06800, Turkey. E-mail: [email protected]
Note. This manuscript was submitted on August 22, 2016; approved on Genetic Optimization of Space Layout and Building
December 9, 2016; published online on March 15, 2017. Discussion period Performance
open until August 15, 2017; separate discussions must be submitted for
individual papers. This paper is part of the Journal of Computing in Civil Genetic algorithm (GA) is a class of evolutionary metaheuristic
Engineering, © ASCE, ISSN 0887-3801. problem-solving techniques based on the principles of biological

© ASCE 04017025-1 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 evolution. The main method of GAs is guided random search that building scales to optimize energy consumption, cost, materials,
gradually improves a population of individuals that are subject to and thermal and lighting behavior. Yi and Malkawi (2009) pro-
fitness pressure. New individuals are generated by recombination posed a method to optimize the energy consumption of complex
operators that exploit the existing genetic material with high building forms by defining hierarchical relationship between
fitness and mutation operators that extend the search toward unex- geometry points. Kämpf and Robinson (2010) developed a meth-
plored areas in the fitness landscape. Genetic algorithm is espe- odology for optimizing building and urban geometric forms for
cially efficient for large, multimodal, and discontinuous search the utilization of solar irradiation using an evolutionary algorithm
spaces, wherein the cost of finding the best solution is too high to minimize energy consumption. Similarly, energy and construc-
and good enough solutions suffice. Genetic algorithms can tackle tion cost are minimized by the hybrid neural network–genetic
both single-objective and multiobjective problems. In single- algorithm approach that Ouarghi and Krarti (2006) developed
objective genetic algorithms (SOGAs), the selection pressure is to optimize office building shape, window-to-wall ratio, glazing
applied through one single fitness measure that is to be minimized type, and insulation. Tuhus-Dubrow and Krarti (2010) developed
or maximized. SOGAs can also be used in multiobjective prob- a simulation-optimization tool and applied it to optimize building
lems by aggregating multiple fitness values in a single-objective shape and envelope features of residential buildings. Building
function (Coello 2002), where each objective value is weighed to orientation, aspect ratio, window type, window-to-wall ratio,
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

balance their satisfaction against others. A multiobjective genetic and wall and roof construction were parameterized in a multiob-
algorithm (MOGA) looks for a Pareto-optimal set of solutions jective optimization model that assists users in green building
while dealing with multiple objectives. Although MOGA shares design by Wang et al. (2005b). Coley and Schukat (2002)
its main principles with SOGA, it differs from it in the selection combined computer-based optimization and human judgment
and retention of the good individuals. Nondominated sorting in a semiautomated model using GA.
genetic algorithm-II (NSGA-II) is a dominance-based MOGA ap-
proach that can effectively balance exploitation and exploration in
its search for optimal design alternatives. This way, it can support Research Gaps in Integral Optimization Approaches
early-design decision making by reaching multiple optimal solu- and Design Strategies
tions and expanding the solution space.
Automated building layout approaches have previously been Early design phases can benefit from genetic optimization to gen-
developed that make use of GAs to optimize spatial geometry erate and evaluate a high number of design alternatives. There are
and topology. Among these works there are a number of heuristics many studies focusing on building shape, opening size and con-
for layout generation. The slicing tree approach uses a binary tree figurations, construction materials, and the design of shadings.
representation to repeatedly subdivide a given building shape by However, the design of spatial layout has not yet been addressed
vertical and horizontal dissections to acquire spaces. Although in performative design tools despite its determinant role in build-
many studies successfully applied this approach in two-dimensional ing performance. To the best of the authors’ knowledge, a com-
(2D) floor plans (Tam 1998; Al-Hakim 2000; Azadivar and Wang putational approach that considers and optimizes spatial layout
2000; Shayan and Chittilappilly 2004; Wu and Appleton 2002; during performative design does not exist. Therefore, an inte-
Honiden 2004; Aiello et al. 2012; Banerjee et al. 2008), it fails grated strategy for the optimization of both spatial layout and
in three-dimensional (3D) layouts, nonsliceable floor plans, and building performance is necessary. Here, computational chal-
nonrectangular spatial and building forms. An alternative heuristic lenges need to be resolved regarding (1) different problem formu-
is to assign spaces to the cells of a matrix representation considered lations to layout and performance optimization, (2) different
as a quadratic assignment problem (Balakrishnan and Cheng 2000; design and solution representations (corresponding to geometry
El-Baz 2004; Mak et al. 1998; Rajasekharan et al. 1998), or space models and chromosome representations) of layout and energy
filling curves (Buscher et al. 2014; Islier 1998; Kochhar et al. 1998; and daylighting optimization, and (3) the need to attend to many
Wang et al. 2005a), or by growing spaces from seeds by rule-based design variables and objectives. Moreover, design processes
heuristics (Inoue et al. 2014). However, these approaches also require the simultaneous exploration of design artifacts, require-
cannot generate 3D and vertically continuous spaces that span ments. and objectives (Van Langen and Brazier 2006), during
multiple floors. A third heuristic packs a number of spatial which a high number of design alternatives need to be systemati-
blocks within a given boundary (Baušys and Pankrašovaite 2005; cally evaluated and benchmarked. Therefore, there is a need for
Dunker et al. 2003; Koenig and Schneider 2012; Lee et al. 2005; design strategies that can systematically explore the solution
Rodrigues et al. 2013). However, the computation time to detect space and provide insight into the solution space for better-
and avoid collision and empty spaces and their limited range of informed decision making.
simplified spatial forms remain an obstacle in their application in
architecture. In conclusion, existing studies mostly deal with 2D
and nonrectangular forms, hindering their effective use in architec- Multiobjective Architectural Design Explorer
tural design. It is necessary to develop a novel heuristic that can
support the design of arbitrarily complex building forms and spatial MADE is a design tool that supports the optimization of building
programs. spatial layout and envelope openings in two submodules (Fig. 1).
Similar to layout design, various metaheuristic methods have MADE-Layout [see Dino (2016) for an extended discussion]
been applied in performance optimization for early-phase build- implements SOGA to find an optimal layout satisfying spatial
ing design. Granadeiro et al. (2013) integrated energy perfor- constraints, while MADE-Opening implements MOGA to deter-
mance assessment in the decision making of building shapes mine the opening sizes that have optimal energy and daylighting
using shape grammars. Lin and Gerber (2014) proposed a deci- performance. The optimal layouts generated by MADE-Layout
sion-support framework that couples parameterization with are provided as input to MADE-Opening. Then, MADE-Opening
MOGA that can generate formal variety while taking energy per- computes a Pareto set of building opening sizes that optimize en-
formance and costs as objectives. Caldas (2008) combined SOGA ergy and daylighting performance. To this end, MADE-Opening
and MOGA with an energy simulation tool applied on various is seamlessly integrated with EnergyPlus, an existing energy

© ASCE 04017025-2 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

Fig. 1. MADE system architecture

simulation tool, to calculate the energy use and daylighting the width (ratioW), depth (ratioD), and height (ratioH) ratios are
performance. generated, and Pi ’s actual widthi , heighti , and depthi are calculated
using these ratios. Then, Pi is assigned to a random voxel in the
building (Cix ; Ciy ; Ciz ) from its center point. Prism Pi is trimmed off
MADE-Layout if it overflows the building. After all the zones are placed in the
building, conflicts regarding zone overlaps and/or empty voxels
MADE-Layout makes use of SOGA to address 3D unequal-area are resolved using two permutation lists, the collision precedence
layout design in three submodules: (1) precedence-based layout list (CPL) and the fill precedence list (FPL). The CPL maintains the
configuration heuristics (P-LCH) is a novel heuristic that generates precedence of each zone in case of space collision. The highest-
layouts without overlaps or empty spaces; (2) the constraint ranking zone in CPL has the precedence to keep the voxel when
checker quantifies layout fitness by means of constraint penalties; multiple zones are assigned to it. The FPL maintains the precedence
and (3) the genetic engine provides feedback to P-LCH to improve of each space to occupy the empty voxels surrounding it. For each
solutions. The user inputs (Table 1, first row) define the require- empty voxel island, the neighboring zone(s) that border these vox-
ments that the layouts need to satisfy based on which layouts els are computed, and the highest ranking in FPL extends itself onto
are constructed (Table 1, second row) and constraint penalties that island. As a result of these two steps, all overlap and empty
are calculated (Table 1, third row). voxel conflicts are resolved. The resulting layout is represented
P-LCH is a novel method that generates valid space layouts as a second 3D matrix (Abuil ) with the same size as Ainit that main-
without zone conflicts or empty areas. As input, P-LCH takes a tains the zones’ indexes. The following generations make use of
discretized building form represented as a 3D Boolean matrix recombination and mutation to generate new individuals.
(Ainit ) and a list of spaces (zones) that are also discretized such that The constraint checker computes layout fitness by penalizing
the total number of the spatial units is equal to the total number of constraints regarding a zone’s form, placement, and topology
voxels. P-LCH assigns zones to voxels as follows. (Table 2). Here, all constraint penalties are normalized to the
In the first generation, for each zone Zi , a rectangular prism Pi (0–1) range to allow intercomparability. A total penalty score
that has the same or approximately the same number of voxels as in fðLÞ for a given layout L is calculated by aggregating the penalties
the brief (NRiv ) is generated. For this, random values over (0, 1) for of all zones, and is minimized toward zero [Eq. (1)]

© ASCE 04017025-3 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 1. User Inputs and Variables of the Zones of the Layouts Generated by MADE-Layout
Input type Description Variable name
User-defined inputs User-defined, discretized building form in a 3D array Ainit
List of zones, and for all Zi in this list:
User-defined number of required voxels for Zi NRiv
User-defined width, height, and length of the bounding box in which Zi is to fit rwi ; rhi ; rli
User-defined façade direction for Zi (north, south, east, or west) facadei
User-defined floor for Zi floi
User-defined maximum number of corners that apply to all zones cormax
Array that maintains topology information (N = neighborhood, S = separation) Atopo
Constraint weights wsize , wdim , wcompact , wjag ,
wconvex , wfacade , wfloor , wneigh , wsep
P-LCH generated information Initial rectangular prism of Zi Pi
The x, y, and z coordinates of the center voxel of Pi Cix , Ciy , Ciz
Width, depth, and height of Pi widthi , depthi , heighti
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

Information used to calculate Actual number of voxels in Zi NAiv

layout constraint penalties Center of gravity point of Zi CoGi
Actual number of corners of Zi cori
Total number of concavities of Zi cnvi
Total number of voxels belonging to Zi that are aligned to facadei V ifac
Total number voxels belonging to Zi that are on floi V iflo
Width, height, and length of the minimum bounding box of Zi awi ; ahi ; ali
Rectangular distance between Zi and Zj rDistðSi ; Sj Þ
Euclidean distance between two points, a and b eDistðPa ; Pb Þ
Vertical distance between voxel V a and flok vDistðV a ; flok Þ
Total number of voxel faces that Si and Sj share faceðSi ; Sj Þ

X
Ns In the genetic engine, MADE-Layout maintains the spatial in-
minfðLÞ ¼ ðwsize Csize þ wdim Cdim þ wcompact Ccompact þ wjag Cjag formation dealt by P-LCH in chromosomes. Two encoding
i¼1
schemes are used: value encoding for the spaces’ rectangular
þ wconvex Cconvex þ wfacade Cfacade þ wfloor Cfloor Þ prisms, and permutation encoding for the two precedence lists.
Atopo ·length In the former, the center point (Cix , Ciy , Ciz ) and dimensions
X
þ ðwneigh Cneigh þ wsep Csep Þ ð1Þ (depthi , heighti ) of each prism Pi is stored (widthi is calculated
j¼1 only when required). In the latter, two permutation lists are used

Table 2. MADE-Layout Constraints

Constraint Constraint
classification name Explanation Mathematical expression
Zone form csize Ensures that a zone contains the c ¼ PN s  h
NAiv NRiv
i
size i¼1 min ;
NRiv NAiv
required number of voxels
cdim Ensures that a zone’s dimensions P s ½maxð0; rwi − awi Þ þ maxð0; rli − ali Þ þ maxð0; rhi − ahi Þ
cdim ¼ Ni¼1
are within user-defined limits in f½maxð0; depthb − awi Þ þ maxð0; widthb − ali Þ þ maxð0; heightb − ahi Þg
three directions
ccompact Ensures the compactness of a zone P s inertiai
c ¼ Ni¼1
by minimizing its moment of inertia compact inertiabuilding
with respect to the rotation axis P iv
inertiai ¼ Nj¼1 eDistðCoGi ; CoGV ij Þ2
defined as the center of gravity 8  
cjag Ensures a zone’s geometric < PN s cori − cormax
max 0; ; cori > cormax
regularity by minimizing its cjag ¼ i¼1
cormax
:
number of corners 0; otherwise
cconvex Ensures the convexity of a zone’s c PN s cnvi
convex ¼ i¼1
form NAiv
Zone cfacade Maximizes a zone’s number of PN s 1
cfacade ¼ i¼1 1 −
placement voxels facing toward a given 1 þ V ifac
building façade
cfloor Maximizes a zone’s number of P s Pm vDistðV ij ; floi Þ
voxels on a user-defined floor cfloor ¼ Ni¼1 j¼0
8 NAiv
Zone cneigh Ensures that two zones are placed P
< N s −1 N sP 1
next to each other by maximizing cneigh ¼ i¼1 j¼iþ1 1 − ; Atopo ði; jÞ ¼ N
topology faceðSi ; Sj Þ þ 1
:
their geometric interface 0; otherwise
8
csep Maximizes the physical distance < PN s −1 PN s 1
csep ¼ i¼1 j¼iþ1 1 − ; Atopo ði; jÞ ¼ S
between two zones rDistðSi ; Sj Þ þ 1
:
0; otherwise
Note: w = user-defined penalty weights.

© ASCE 04017025-4 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 3. Simulation Settings, Decision Variables, and Objective Functions
Variable
category Settings Variable name Unit
Simulation Simulation Run period Dates
parameters settings
Weather data —
Material Thermal conductivity (λ) W=mK
properties
Thickness (th) m North
Solar transmittance at normal —
incidence (T sol ) only for windows
Visible transmittance at normal —
incidence (T vis ) only for windows Perspective view of Z i
Zone information Thermal loads and schedules W=m2
(for each zone Zi ) (Qpeople , Qlight , Qequip )
wwr N
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

Emin
Zi and setpoint schedules lx Zi wwr W
Zi
Heating and cooling setpoint °C
schedules
HVAC (standard Zone controls, equipment —
for all buildings) configurations, ideal loads system
component
North side West side
Decision Window-to-wall ratio (wwrNZi , —
variables wwrSZi , wwrEZi , wwrW Zi ) wwr S
Zi wwr E
Zi
Objective BTEU kWh
functions
BDA h

to maintain the precedence information of the spaces. When a South side East side
genotype is to be mapped to its phenotype (a valid 3D layout),
the P-LCH algorithm is used. Fig. 2. Window placement example of a zone with 12 voxels (Zi )
Elitism, crossover, and mutation operations are implemented as
genetic operators. In elitist selection, the best individuals are copied
to the new generation with a rate of Pe . The remaining individuals
are selected for crossover with a rate of Pc . Uniform crossover is for each voxel with user-specified dimensions (widthi , depthi ,
applied on the chromosome’s value-encoded part, while partial heighti ). An energy model is built in OpenStudio SDK, where each
mapping crossover (PMX) is used for the two ordered lists. zone corresponds to a thermal zone. Following, the parameters nec-
Afterward, a mutation is applied to individuals with a rate of essary for simulation such as materials, internal loads, schedules,
Pm . The mutation replaces alleles with random values on the chro- and illuminance setpoints are entered by the user (Table 3). If there
mosome’s value-encoded part, or switches two random gene posi- are neighboring buildings or other obstacles that may influence the
tions with each other on the ordered lists. MADE-Layout energy calculations, they also need to be added in the model by
terminates when the highest fitness value remains the same for a specifying their surface reflectivity and geometry in relation to
user-determined number of successive generations (N ni ). the building. The wwr values are determined randomly for the first
generation. In the following generations, the genetic operators
compute these values.
MADE-Opening aims to minimize BTEU and maximize BDA.
MADE-Opening
BTEU aggregates the building heating, cooling, and lighting en-
MADE-Opening supports the optimization of building energy and ergy use [Eq. (2)]. BDA is calculated in two steps. First, the day-
daylighting performance using NSGA-II as a multiobjective opti- z ) for each zone Zi for a given hour h is
light illuminance value (Ecalc
mizer. The zones’ window-to-wall ratios (wwr) are selected as de- calculated by EnergyPlus. Then MADE-Opening calculates day-
sign variables, and building total energy use (BTEU) and building light illuminance setpoint satisfaction (DISS) using an asymmetric
daylight autonomy (BDA) are selected as the objectives of MOGA. triangular fuzzy membership function, as an indication of how
Both BTEU and BDA are calculated at once by the EnergyPlus close Ecalc
z is to Zi ’s user-defined daylight illuminance setpoint
energy simulation software, which is seamlessly coupled to (Emin
z ). In this function, the lower and upper limits of deviation
MADE-Opening through OpenStudio SDK, an open-source frame- (ElowB
z and EhighB
z ) can be stretched by the user to specify the al-
work that supports tool development by providing access to model lowable daylighting values. The DISS value then is weighed by the
object attributes for the programmatic generation of energy models number of voxels Zi contains (NAiv ). Finally, the eventual BDA
(Weaver et al. 2012). Finally, MADE-Opening presents to the value is calculated by aggregating the DISS values for all the hours
users a Pareto optimal set of nondominated solutions for further of the given design days for all zones [Eq. (3)]
exploration.
minBTEU ¼ Qheating þ Qcooling þ Qlighting ð2Þ
Fitness Evaluation: Energy Modeling and Simulation
MADE-Opening takes as input a layout solution (Abuil ) that the X
t X
k
user generates in MADE-Layout. First Abuil is converted from a maxBDA ¼ DISShz NAiv
3D matrix into a 3D form by instantiating a rectangular prism z¼1 h¼0

© ASCE 04017025-5 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Chromosome Representation
MADE-Opening’s chromosomes maintain the wwr values of
the zones. The following rules determine opening size and place-
ment on the building. Each voxel belonging to Zi initially has to
Fig. 3. MADE-Opening chromosome representation have exactly one window on all of its four exterior walls, located
on the center of the voxel face (Fig. 2). During genotype-to-
phenotype mapping, if a side of a voxel is detected as facing
the building interior, it is omitted. Of all voxels of Zi , the walls
8
> facing the same direction need to have the same wwr value, but
>
> 0; Ecalc
z ∈ ½0; ElowB
z Þ
> Ecalc
> z −Ez
different directions have different wwr values. MADE-Opening
< min
min
lowB þ 1; Ez ∈ ½Ez ; Emin z Þ
calc lowB
Ez −Ez chromosomes maintain all the zones’ wwr information in all four
DISShz ¼ Ez −Ez
calc min ð3Þ
>
> highB þ 1; Ecalc ∈ ½Emin highB
z ; Ez  directions (Fig. 3). Similar to MADE-Layout, value encoding is
>
> Emin
z −Ez
z
>
: 0; used in the chromosome, where the wwr values can vary between
Ecalc
z ∈ ½EhighB
z ; ∞Þ
(0 and 1).
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

Form

Northeast view Southwest view

Form

Northeast view Southwest view

Form

Northeast view Southwest view

Form

Northeast view Southwest view

Fig. 4. Discretized building forms, northeast and southwest views

© ASCE 04017025-6 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 4. Building Zones and Variables for Constraint Calculation
Zone Floor area (m2 ) NRiv facadei floi rwi rhi rli Neighborhood Separation
Reading 1, Z1 1,600 16 — — 5 5 4 — —
Reading 2, Z2 1,100 11 — — 5 5 4 Z4 —
Books, Z3 1,100 11 — — 5 5 4 Z4 —
Administration, Z4 800 8 — — 5 5 3 Z2 , Z3 , Z 5 , Z6 —
Café 1, Z5 600 6 South 3 5 5 1 Z4 , Z 7 Z8
Working, Z6 1,800 18 — — 5 5 5 Z4 —
Conference room, Z7 100 1 North — 1 1 1 Z5 —
Café 2, Z8 100 1 — 0 1 1 1 — Z5

Selection to validate the MADE by demonstrating its use through a design

exploration strategy that aims to carry out comparative Pareto-
NSGA-II, a Pareto-based elitist approach that applies nondomi-
based studies. The strategy presented here aims to use MADE
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

nated sorting in its selection process (Deb et al. 2002), is imple-

to (1) generate building alternatives with different forms, orienta-
mented. In NSGA-II, two ranking selection operations aim to
tions, and layouts; (2) maintain not one but a number of design
balance exploitation and exploration. For exploitation, NSGA-II
alternatives at the same time; and (3) comparatively evaluate
demonstrates a higher bias toward the selection of Pareto-dominant
and benchmark these alternatives based on their fitness values. This
solutions. In general, in a population Pi of size N, a solution S1 is
strategy can set an example of similar design processes that makes
said to dominate solution S2 if it is no worse than S2 in all objec-
effective use of MADE by systematically exploring the solution
tives and if S1 is better than S2 in at least one objective. All indi-
space for better-informed decision making.
viduals in the current population are assigned a nondomination
rank (Si·rank ) that counts the number of other individuals that domi-
nate them, which is an indication of how well that individual per- MADE-Layout
forms. For exploration, clustering needs to be eliminated to achieve
a uniform distribution and preserve solution diversity. Accordingly, The building being studied is located in Ankara, Turkey. The ar-
all individuals are also assigned a crowding distance (SCn chitectural program requires a floor area of 7,200 m2 and consists
i·distance )
value for all objectives Cn that indicate their neighborhood density. of eight thermal zones. MADE-Layout takes as input a number of
Crowding distance is represented by Eq. (4), where CSi library building forms (Fig. 4) and the zones to be allocated in the
n is the ob-
jective value of solution Si , and Cnmax and Cnmin are the maximum buildings provided by the architectural firm, as well as their con-
and minimum values of Cn straint settings (Table 4), weights (Table 5), and GA parameters
(Table 6). All forms and zones are discretized into the same number
of units (72 units). The maximum corner constraint (cormax ) is
CSi−1
n − CSiþ1
n
SCn
i·distance ¼ ð4Þ set to 12.
Cn − Cnmin
max
For each form, 10 MADE-Layout simulations were performed.
Each form’s highest-ranking layout is presented in Fig. 5. In
Next, individuals are selected for mating using binary tourna- addition, the four highest-ranking layouts of Form B are presented
ment selection. When two chromosomes with unequal ranks are in Fig. 6. The constraint penalties and fitness values vary from
competing, the one with the lower nondomination rank is selected. layout to layout. In conclusion, form compactness and high voxel
Otherwise, the one with the higher crowding distance is selected. connectivity positively contributed to higher-ranking individuals.
Because two individuals are selected for mating, uniform crossover
is applied on chromosomes. The parent that will provide its allele to
Table 5. Constraint Weights
the offspring will be determined by a probability value Po of 70%
for either parent. Additionally, an allele of an individual can be se- Constraint Weight
lected for mutation with a probability value of Pm . In this case, the csize 200
existing value is replaced by a new random value. As a result, a cdim 250
temporary population Qi is generated by these new individuals. ccompact 10
Values Pi and Qi are merged into a population of size 2N, and cjag 750
the best individuals are transferred for the next population Piþ1 . cconvex 10
This feedback loop continues until the termination condition is cfacade 20
met, which is reaching the user-defined number of generations. cfloor 80
cneigh 30
Finally, the nondominated solutions (Pareto front) of the final pop-
csep 1
ulation are presented to the user.

Case Study Table 6. Genetic Algorithm Parameter Values

MADE is evaluated through a proof-of-concept software prototype GA parameter Value

implemented in Java. To automate the buildup of energy models Pc 0.9
and their simulations, MADE-Opening is integrated with Energy- Pe 0.1
Plus 8.2 through OpenStudio SDK. To test MADE, the design spec- Pm 0.1% × nonimproving generation count
ifications of a library building that is being designed by an Pr 0.1% × nonimproving generation count
architectural firm are readily used. The form, spatial, and perfor- Np 1,000 individuals
Nt 500 nonimproving generations
mance information are also provided by this firm. This section aims

© ASCE 04017025-7 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Layout A6
Fitness 23.57195649
csize 0.02178030
cdim 0.0
ccompact 0.39085365
cjag 0.0
cconvex 0.0
cfacade 0.32142857
cfloor 0.0
cneigh 0.27777777
csep 0.0
Northeast view Southwest view
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

Layout B5
Fitness 27.69714164
csize 0.0
cdim 0.0
ccompact 0.14925498
cjag 0.00637755
cconvex 0.0
cfacade 0.35
cfloor 0.0
cneigh 0.460714286
csep 0.6
Northeast view Southwest view

Layout C2
Fitness 26.73771788
csize 0.02178030
cdim 0.0
ccompact 0.21868312
cjag 0.00347222
cconvex 0.0
cfacade 0.35
cfloor 0.0
cneigh 0.32738095
csep 0.0
Northeast view Southwest view

Layout D4
Fitness 30.51772142
csize 0.03645833
cdim 0.0
ccompact 0.17941780
cjag 0.0019531
cconvex 0.0
cfacade 0.32142857
cfloor 0.0
cneigh 0.43333333
csep 0.0
Northeast view Southwest view

Z1 Z2 Z3 Z4
Z5 Z6 Z7 Z8

Fig. 5. Highest ranking layout of each building form, the final fitness and raw penalty values; northeast and southwest views

© ASCE 04017025-8 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Layout B1
Fitness 33.08356789
csize 0.0
cdim 0.0
ccompact 0.18104386
cjag 0.00765306
cconvex 0.0
cfacade 0.35
cfloor 0.0
cneigh 0.60000000
Northeast view Southwest view csep 0.533333333
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

Layout B3
Fitness 34.59124096
csize 0.04848710
cdim 0.0
ccompact 0.19506230
cjag 0.00510204
cconvex 0.0
cfacade 0.32142857
cfloor 0.0
cneigh 0.39404762
Northeast view Southwest view csep 0.86666667

Layout B5
Fitness 27.69714164
csize 0.0
cdim 0.0
ccompact 0.14925498
cjag 0.00637755
cconvex 0.0
cfacade 0.35
cfloor 0.0
cneigh 0.46071429
Northeast view Southwest view csep 0.6

Layout B7
Fitness 28.92860827
csize 0.0
cdim 0.0
ccompact 0.15173838
cjag 0.00765306
cconvex 0.0
cfacade 0.37500000
cfloor 0.0
cneigh 0.45238095
Northeast view Southwest view csep 0.6

Z1 Z2 Z3 Z4
Z5 Z6 Z7 Z8

Fig. 6. Form B’s four highest ranking layouts, their final fitness and raw penalty values; northeast and southwest views

© ASCE 04017025-9 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 7. NSGA-II and BDA Fuzzy Membership Variables Table 9. Glazing Material
NSGA-II parameters Value Property Value
Crossover rate, Pc 0.8 Thickness 0.006
Mutation rate, Pm 0.1 Conductivity (W=mK) 0.9
Population size, N p 40 individuals Visible transmittance at normal incidence 0.44
Number of iterations, N t 40 generations Solar transmittance at normal incidence 0.3311
Lower limit of DISS 0.6Emin
z
membership function, ElowB
z
Upper limit of DISS 1.5Emin
z comparison of building forms, layouts, and orientation. For each,
membership function, EhighB
z
fitness values are calculated and the Pareto sets are computed and
visualized. The resulting Pareto sets provide a rich design space for
design exploration, wherein alternative designs can be bench-
An extended discussion on variable dependencies, constraint-based marked against each other. As a quantitative measure to compare
design exploration, population convergence, and genetic parame- Pareto fronts, the HV metric is used. Hypervolume measures the
ters can be found in Dino (2016). Finally, the highest-scoring lay-
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

volume of the region populated by the nondominated set with re-

out of each form (Fig. 5) and the four highest-ranking layouts of spect to a reference point (Zitzler and Thiele 1999). Because HV
Form B (Fig. 6) are selected as input into MADE-Opening. computes the size of the space dominated by the Pareto front, cov-
ering a maximal HV is an indicator of both good convergence
(closeness to the utopian point) and diversity (a long distribution
MADE-Opening
along the Pareto front with a variety of solutions). In this study,
MADE-Opening operates on the layouts previously generated by HV is used to benchmark multiple building instances through their
MADE-Layout using NSGA-II (settings presented in Table 7) and Pareto fronts. Hypervolume is defined in Eq. (5), where HVb is the
EnergyPlus. Simulation-specific information needs to be provided HV of building b, PFb is the Pareto front, RP is the reference point,
by the users in order to calculate the fitness values. The case study n is the number of individuals on PFb , and vi is the volume of the
uses the following setup: space defined between RP and the individual i. Parameter RP is
• EnergyPlus simulation is run only for 4 days that are represen- defined as the point in the objective space that is dominated by
tative of each season regarding solar irradiation and outside tem- all nondominated solutions
perature (January 21, April 21, July 21, and October 21). These !
data are obtained from the Climate Design Data 2009 ASHRAE [n
HVb ðPFb ; RPÞ ¼ volume vi ð5Þ
Handbook (ASHRAE 2009) as encoded as weather files in
i¼0
EnergyPlus (WMO 171280, Ankara Esenboga).
• To build the 3D geometry, each voxel is converted into a rec-
tangular prism of 10 × 10 × 5 m. The prism surfaces that face Form Exploration
the exterior are instantiated as walls, roofs, or floors with dif- In this phase, the highest-ranking layouts generated by MADE-
ferent materials (Table 8). The glazing material for the windows Layout (A6, B5, C2, and D4) are studied. MADE-Opening is
is given in Table 9. Each zone is instantiated as a thermal zone run for each layout. Their Pareto front is plotted (Fig. 7) and
with different internal gains and setpoints (Table 10). HV values are calculated (Table 11). The results show that B5
• EnergyPlus’s ideal HVAC system is used, which calculates and C2 have the highest HV values with similar Pareto fronts.
loads without the need to model a full HVAC. The system mixes Layout A6 performed well in energy but failed in daylighting.
zone air with the specified amount of outdoor air, adding or re- Conversely, D4 scored high daylighting values but could not reduce
moving heat and moisture at 100% efficiency (U.S. DOE 2010). energy use as much as the other layouts. A combined analysis re-
For space heating and cooling, dual setpoint thermostats are garding the layout performance (from MADE-Layout) and energy
used. and daylighting performance (HV values) reveals that they are
• To eliminate excessive daylight illuminance, controllable win- negatively correlated, such that best layout fitness values score
dow shading elements on the south-facing, east-facing, and the lowest HV as in A6. This means that, contrary to the common
west-facing windows that are dynamically adjusted to meet assumption, compact forms A6 and D4 did not perform as well as
the user-defined daylight illuminance setpoint (Emin z ) are used fragmented forms B5 and C2. This points to a trade-off between the
at each zone’s daylight reference point. Other shading elements building form, energy, and daylight performance, calling for some
can also be selected. degree of compromise in the final selection.
Using the preceding setup, MADE-Opening is tested to carry The second analysis involves the comparison of the individuals’
out different exploratory studies regarding the generation and wwr values. One solution per Pareto front on the break point of the

Table 8. Materials, Listed from Outside to Inside

Envelope material Name Roughness Thickness (m) Conductivity (W=mK) Density (kg=m3 ) Specific heat (J=kg · K)
Exterior wall Stucco Smooth 0.0253 0.6918 1,858 837
Concrete Medium-rough 0.2033 1.7296 2,243 837
Wall insulation Medium-rough 0.0794 0.0432 91 837
Gypsum Smooth 0.0127 0.16 784.9 830
Roof Roof membrane Very rough 0.0095 0.16 1,121.29 1,460
Roof insulation Medium-rough 0.2105 0.049 265 836
Metal decking Medium-rough 0.0015 45.006 7,680 418.4
Floor and ceiling Concrete Medium-rough 0.2033016 1.729577 2,242.585 836.8
Interior wall Concrete Rough 0.2033016 0.5707605 608.7016 836.8

© ASCE 04017025-10 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 10. Zones’ Thermal and Illumination Requirements
Illumination Maximum daylight HVAC heating HVAC cooling
Zone People (per m2 ) Lights (W=m2 ) Electric equipment (W=m2 ) setpoint (EZi ) (lx) glare index (DGI) setpoints (°C) setpoints (°C)
Z1 0.056511 10 10.76 500 22 22 25
Z2 0.11 10 10.76 500 22 22 25
Z3 0.005 10 0.58 150 22 20 26
Z4 0.056511 10 10.76 400 22 22 25
Z5 0.29 10 6.4 250 22 23 26
Z6 0.056511 10 10.76 500 22 22 25
Z7 0.056511 10 10.76 150 22 22 25
Z8 0.29 10 6.4 250 22 23 26
Note: The fraction to be applied to the internal gains throughout the day are 12 to 7 a.m. = 0, 7 to 8 a.m. = 0.25, 8 a.m. to 12 p.m. = 1, 12 to 2 p.m. = 0.75,
2 to 6 p.m. = 1, 6 to 7 p.m. = 0.5, 7 to 8 p.m. = 0.25, 8 p.m. to 12 a.m. = 0.
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

that have similar BTEU values, B3-i26 and B7-i20, are selected
(visualized in Fig. 10, together with B1 and B5). Their objective
values (Table 14) reveal that Z2 and Z4 of B3-i26 cannot satisfy
their high illumination requirements due to their horizontal depth
and north exposure. As a result, it can be concluded that an optimal
solution (1) favors shallow zones and (2) avoids the northern
exposure of zones with higher illumination needs.

Orientation Exploration
For orientation studies, B5, a fragmented form with an aspect ratio
of 7:4, is selected and its four different azimuth angles, B5-α0 ,
B5-α90 , B5-α180 , and B5-α270 , are simulated. Because façade ori-
entation was a part of in MADE-Layout’s fitness function (cfacade ),
the layout fitness value is first recalculated through MADE-Layout,
such that it corrects the cfacade penalty of the rotated layout. Then,
their Pareto fronts are computed by MADE-Opening and HV val-
ues are calculated (Fig. 11 and Table 15). The east-west oriented
Fig. 7. Pareto set of A6, B5, C2, and D6; the individuals on the layout B5-α0 could minimize BTEU slightly better than B5-α90
3,750-kWh line are visualized in Fig. 8 and B5-α270 , but is very similar in BDA. Layout B5-α180 , on
the other hand, too quickly reaches a low plateau regarding day-
lighting. Again, one individual on the Pareto front of each orien-
tation (B5-α0 -45, B5-α90 -4, B5-α180 -22, and B5-α270 -17) is
Table 11. HV and Layout Fitness Values of A6, B5, C2, and D4 selected and its energy and daylighting values are plotted
(Table 16). It is observed that B5-α180 -22’s daylighting satisfaction
Building and of Z4 and Z6 are significantly lower compared with the other zones.
layout A6 B5 C2 D4
As previously, the north orientation and the horizontal depth of
HV 1.0823 × 106 1.8807 × 106 1.6592 × 106 1.2589 × 106 these zones hinder their reference illumination points reaching
Layout fitness 23.57 27.69 26.73 30.51 the target illuminated levels.
(previously
calculated by Higher-Resolution Design Explorations
MADE-Layout) To evaluate a high number of individuals rapidly, MADE-Opening
deals with low-resolution energy models and 4 design days for sim-
ulation. As a result, yearly simulations are postponed until the sol-
3,750-kWh BTEU line is visualized (Fig. 8). The layouts’ average ution space is narrowed to a limited number of alternatives. In this
wwr values show that MADE-Opening consistently maximized section, the results of detailed simulations for the whole year for the
the south and minimized the north wwr values (Table 12). The best-performing building instance, B5-α0 -45, are presented. As a
exception to this rule are zones with a single north-facing window result the yearly, monthly, or daily energy use and daylighting per-
that try to increase its wwr to receive sufficient daylight. Although formance can be analyzed for each zone (Fig. 12 and Table 17).
western solar exposure is generally not desired due to extreme heat These analyses can calculate the actual building performance (as
gain, MADE generated unusually large western windows due to the opposed to simplified simulations that consider a limited number
controllable shades that limit the undesired heat gain. of days) and allow benchmarking between other building instances
with the same use and size.
Layout Exploration
In this phase, Form B’s four highest-ranking layouts are studied
(B1, B3, B5, and B7). Their Pareto sets are plotted (Fig. 9) and Discussion
their HV values are calculated (Table 13). The HV results suggest
that B7 performs the best, representing a good distribution over the
objective space and a close proximity to the true Pareto front. A Size of the Variable Space and Objective Space
noticeable gap between B3 and the other layouts’ HV values is ob- MADE addresses a dense objective space (satisfying a number
served. To understand the reason, individuals from B3 and B7 of spatial constraints for layout design and ensuring energy and

© ASCE 04017025-11 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 A6.i50
Objective 1
BTEU (kWh): 3767.83
Qelectric: 1371.74
Qheating: 1303.87
Qcooling: 1092.23
Objective 2
BDA (hours): 2121.23

Northeast view Southwest view

Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

B5.i3
Objective 1
BTEU (kWh): 3759.54
Qelectric: 1319.59
Qheating: 1468.46
Qcooling: 971.49
Objective 2
BDA (hours): 2447.07

Northeast view Southwest view

C2.i38
Objective 1
BTEU (kWh): 3734.46
Qelectric: 1335.32
Qheating: 1369.10
Qcooling: 1030.04
Objective 2
BDA (hours): 2455.36
Northeast view Southwest view

D4.i27
Objective 1
BTEU (kWh): 3764.73
Qelectric: 1407.40
Qheating: 1081.98
Qcooling: 1275.35
Objective 2
BDA (hours): 2092.33

Northeast view Southwest view

Z1 Z2 Z3 Z4
Z5 Z6 Z7 Z8

Fig. 8. Objective values of the layouts selected in Fig. 7; northeast and southwest views

© ASCE 04017025-12 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 12. wwr Values of the Selected Building Instances in Fig. 7 However, they demonstrate different convergence characteristics
Building Gross North East South West due the different characteristics of their fitness landscapes. The
layout problem has a multimodal and rugged fitness landscape.
A6.i50 29.95 2.24 51.59 48.42 36.05
Ruggedness, in general, is related to a low correlation between ad-
B5.i3 29.50 4.29 16.06 51.51 49.72
C2.i38 28.77 5.06 29.49 47.33 33.20
jacent points in a search space, which results in a dramatic change
D4.i27 23.48 3.13 20.59 35.05 35.14 in fitness even with a slight change in design variables. MADE-
Average 27.92 3.68 29.43 45.57 38.52 Layout’s CPL and FPL can have such drastic effects on zones’ form
and topology constraints, and therefore can challenge convergence.
As a solution, MADE-Layout’s layouts were discretized into a 3D
matrix representation, thereby shrinking the search space and
allowing the evaluation of a high number of alternatives in a short
amount of time. The fitness landscape of MADE-Opening is
smoother and more continuous in comparison with MADE-Layout.
On the other hand, calculation of fitness through energy simulation
is computationally expensive and therefore a significantly lower
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

number of individuals can be evaluated. The disproportion between

the average duration of fitness calculation in both MADE modules
(Table 18) represents a trade-off between speed and accuracy. If
combined with MADE-Opening, MADE-Layout’s convergence
could be challenged as a result of processing too few individuals.
The disparity between the convergence behaviors of MADE-
Opening and MADE-Layout additionally justifies their separation
into a two-step procedure.

Optimal Building Designs

The Pareto-based studies in the case study show that compact forms
Fig. 9. Pareto set of B1, B3, B5, and B7; the individuals on the with high layout scores perform poorer in the combined energy and
3,500-kWh line are visualized in Fig. 10 daylighting analyses. More fragmented forms with a larger surface
area could provide more daylighting while maintaining the same
energy performance as others. When the effect of different layouts
of the same form is studied, it is seen that a zone’s horizontal depth
Table 13. HV and Layout Fitness Values of B1, B3, B5, and B7 and orientation have a huge impact on daylighting, and can dras-
Building tically worsen daylight performance especially if that zone has high
and layout B1 B3 B5 B7 illumination requirements. As a result, MADE-Opening favored
HV 1.7492 × 106 1.2983 × 106 2.1284 × 106 2.5669 × 106 layouts with shallower zone forms. The study of different azimuth
Layout fitness 33.08 34.59 27.69 28.92 angles of the same form showed that the east-west orientation is
(previously preferred to north-west orientation as expected.
calculated by
MADE-
Layout) User Interaction and Postprocessing of Designs
The effective use of MADE requires a systematic strategy for design
exploration, as demonstrated in the preceding case study, that in-
daylight performance) that necessitate different solution represen- volves multiple simulations experimenting with different decision
tations. In the development of MADE, two different approaches variables and their comparative evaluation. Tools such as MADE
were followed: (1) aggregating multiple constraints into one automate building design based on objective measures regarding
single-objective function (MADE-Layout); and (2) multiobjective building performance. This might run the risk of unquantifiable cri-
Pareto-based approaches that optimize multiple objectives at once teria, be disregarded such as functional or visual qualities. In this
(MADE-Opening). In this research, MADE took one step further in case, the design formulation of MADE might conflict with users’
combining these approaches into one integral model that serially aesthetical preferences. To involve users in the design process, user
optimizes many objectives in two steps. This also necessitated that interaction is made possible in the selection of the layouts that
decision variables are also distributed among the two modules. MADE-Layout generated into MADE-Opening. Generally, it is
Such variables are formulated either as design constraints and accepted that divergent exploration and processing many design al-
weights, simulation settings, or the initial building form. This ap- ternatives simultaneously is an indication of creativity (Jones 1992).
proach can be generalized to a wider range of design problems in In this process, therefore, the selection of a large number of layout
which design problems require different solution formulations, but alternatives for MADE-Opening can increase the possibility of
a way to incorporate these formulations exists. As such, the prob- reaching an optimal overall design. MADE’s two-step workflow
lem with optimizing too many objectives with a high number of thus broadens and narrows the solution space by providing a
decision variables can be thoroughly tackled. balanced mix of exploration and exploitation.
MADE-Opening determines the opening sizes based on the
building’s voxel representation, which does not reflect a typical
Fitness Landscapes and Algorithm Convergence
building representation. Moreover, the computed wwr values
The case study shows that both MADE-Layout and MADE- represent the optimal condition based on BTEU and BDA. When
Opening converged to either a good solution or to a Pareto front. esthetical concerns are considered, similar wwr values on one

© ASCE 04017025-13 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 B1.i52
Objective 1
BTEU (kWh): 3510.70
Qelectric: 1343.27
Qheating: 1013.17
Qcooling: 1154.26
Objective 2
BDA (hours): 2285.27
Northeast view Southwest view
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

B3.i26
Objective 1
BTEU (kWh): 3499.59
Qelectric: 1429.16
Qheating: 1092.19
Qcooling: 978.24
Objective 2
BDA (hours): 1948.69

Northeast view Southwest view

B5.i45
Objective 1
BTEU (kWh): 3507.36
Qelectric: 1334.46
Qheating: 1248.60
Qcooling: 924.30
Objective 2
BDA (hours): 2378.09

Northeast view Southwest view

B7.i20
Objective 1
BTEU (kWh): 3474.00
Qelectric: 1285.98
Qheating: 1159.41
Qcooling: 1028.62
Objective 2
BDA (hours): 2679.70

Northeast view Southwest view

Z1 Z2 Z3 Z4
Z5 Z6 Z7 Z8

Fig. 10. Objective values of the instances selected in Fig. 9; northeast and southwest views

© ASCE 04017025-14 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 14. Objective Values of the Zones of the Individuals Presented in Fig. 11
BTEU (kWh) BDA (h)
Zone B1-i52 B3-i26 B5-i45 B7-i20 B1-i52 B3-i26 B5-i45 B7-i20
Z1 720.17 797.74 693.69 741.63 602.84 374.83 250.17 574.42
Z2 351.17 639.76 416.52 420.90 257.63 287.07 335.62 399.94
Z3 260.65 301.69 341.25 277.25 173.19 343.35 418.00 435.57
Z4 435.40 335.76 423.79 419.83 301.77 120.60 277.81 306.59
Z5 584.01 285.64 554.87 545.03 221.00 228.00 228.52 239.26
Z6 1035.86 993.82 969.01 989.14 676.25 590.89 624.31 671.90
Z7 54.83 77.32 53.84 49.47 29.94 1.14 2.30 14.90
Z8 68.62 67.87 54.40 30.75 22.66 2.80 1.02 37.13
Total 3,510.70 3,499.59 3,507.36 3,474.00 2,285.27 1,948.69 2,137.73 2,679.70
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

façade might be preferable. In this case, the generated solutions can

be manually modified to explore different opening sizes and forms.

Conclusion

MADE addresses the need for early-phase design tools that support
performance-based design. Building form, layout, and envelope
openings have a huge impact on energy and daylighting perfor-
mance. Their combined optimization in early-design processes
can lead to well-informed decisions toward higher performance.
Layout and energy and daylighting optimization are characterized
by different convergence behavior and fitness landscapes. Corre-
spondingly, MADE-Layout and MADE-Opening should be opera-
tionalized using different design exploration strategies. The former
requires multiple separate runs to reach different optimal solutions,
while the latter computes a Pareto frontier on which users can nego-
tiate and agree on a trade-off solution. The presented case study
Fig. 11. Pareto front for four azimuth angles of layout B5
demonstrates the feasibility of MADE in sequentially optimizing
multiple design objectives that cannot be tackled at once. More-
over, genetic algorithms coupled with simulation tools proved to
be able to facilitate both design exploration within a rich design
space and exploitation by converging to the Pareto frontier.
The simulation-based evaluation function of MADE-Opening
Table 15. HV and Recalculated Layout Fitness Values of Four Azimuth necessitates energy models early during design. While there is extra
Angles
burden in the buildup of these models, it also contributes to a con-
Building and tinuous evaluation supported by simulation. The models initiated
layout B5-α0 B5-α90 B5-α180 B5-α270 by MADE-Opening can be gradually developed further as detailed
HV 2.3114 × 106 2.5294 × 106 2.1117 × 106 2.3964 × 106 design information is made available by time. While the energy and
Layout fitness 27.69 27.98 28.17 27.98 daylighting models can be sustained until the end of design, it is
(recalculated also necessary to integrate other performance analysis tools that
in MADE- were not addressed in this research into the process. To this
Layout) end, future work should consider strategies for data interoperability.
Building information models such as inductry foundation classes

Table 16. Objective Values of the Zones of B5’s Azimuth Angles

BTEU (kWh) BDA (h)
Zone B5-α0 -45 B5-α90 -4 B5-α180 -22 B5-α270 -11 B5-α0 -45 B5-α90 -4 B5-α180 -22 B5-α270 -11
Z1 693.69 757.94 677.84 647.15 492.70 575.68 569.33 571.44
Z2 416.52 425.42 338.20 401.58 291.85 339.57 343.59 318.92
Z3 341.25 284.18 302.28 299.99 439.99 416.89 359.14 380.52
Z4 423.79 435.89 497.67 448.84 306.90 268.37 145.85 261.99
Z5 554.87 569.06 575.40 569.94 233.96 234.88 219.84 216.18
Z6 969.01 935.47 1,009.75 1,014.98 663.38 614.21 497.44 618.17
Z7 53.84 36.37 41.97 36.81 2.30 39.93 41.34 37.20
Z8 54.40 62.30 69.63 68.15 15.99 33.93 40.83 38.00
Total 3,507.36 3,506.64 3,512.72 3,487.43 2,447.07 2,523.46 2,217.36 2,442.43

© ASCE 04017025-15 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

Fig. 12. Yearly heating, cooling, and lighting energy use and daylight autonomy of B5-α0 -45

Table 17. Objective Values—Yearly Heating, Cooling, and Lighting Energy Use and Daylight Autonomy—of B5-α0 -45
Objective values Z1 Z2 Z3 Z4 Z5 Z6 Z7 Z8 Total
Heating (kWh) 53,583.41 18,022.07 23,825.95 22,223.64 20,328.7 56,703.39 3,047.819 1,451.088 19,9186.1
Cooling (kWh) 18,774.06 9,147.469 33.24476 20,103.12 36,181.13 47,168.82 56.44864 2,650.728 13,4115
Lighting (kWh) 27,270.72 19,042.29 18,831.8 12,920.67 9,633.258 29,240.41 1,593.495 2,553.945 12,1086.6
BTEU (kWh) 99,628.19 46,211.83 42,690.99 55,247.43 66,143.08 133,112.6 4,697.762 6,655.761 454,387.7
BDA (h) 2,661.167 2,830.333 3,068 3,392 3,451.5 3,371.167 3,447.833 0 22,222

© ASCE 04017025-16 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Table 18. Convergence Behavior of B5-α0 -45 Selected Papers of the Int. Annual Conf. of the German Operations
Research Society (GOR), S. Helber, et al., eds., Springer, Cham,
Np Cag Cas Average time for
Switzerland, 109–114.
MADE (individuals (generations [convergence fitness evaluation
module per generation) processed) time (s)] (s/individual) Caldas, L. (2008). “Generation of energy-efficient architecture solutions
applying GENE_ARCH: An evolution-based generative design sys-
MADE- 1,000 629 92.6 0.00014721 tem.” Adv. Eng. Inf., 22(1), 59–70.
Layout Coello, C. A. C. (2002). “Theoretical and numerical constraint-handling
MADE- 30 40 (fixed) 6,049 4.95 techniques used with evolutionary algorithms: A survey of the state
Opening of the art.” Comput. Methods Appl. Mech. Eng., 191(11), 1245–1287.
Coley, D. A., and Schukat, S. (2002). “Low-energy design: Combining
computer-based optimisation and human judgement.” Build. Environ.,
(IFC) and green building XML (gbXML) offer a viable solution to 37(12), 1241–1247.
information exchange and tool integration. Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002). “A fast and
Future development can additionally take into account other de- elitist multiobjective genetic algorithm: NSGA-II.” IEEE Trans. Evol.
sign variables regarding building envelope that have an effect on Comput., 6(2), 182–197.
Dino, I. G. (2016). “An evolutionary approach for 3D architectural space
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

energy and daylighting. Material types, both opaque and glazing,

layout design exploration.” Autom. Constr., 69, 131–150.
used in the building envelope can be varied, or different types of
Dunker, T., Radons, G., and Westkämper, E. (2003). “A coevolutionary
shading devices can be considered. Moreover, other objectives can
algorithm for a facility layout problem.” Int. J. Prod. Res., 41(15),
be integrated into the fitness function in addition to energy and day- 3479–3500.
lighting performance. For instance, the minimization of the initial El-Baz, M. A. (2004). “A genetic algorithm for facility layout problems
and lifecycle cost of the building, two objectives that are usually in of different manufacturing environments.” Comput. Ind. Eng., 47(2),
conflict with each other, is a key target in building design, and are 233–246.
good candidates as design objectives. Likewise, initial and lifecycle EnergyPlus [Computer software]. U.S. Dept. of Energy, Washington, DC.
carbon emissions are determinant in the building’s environmental Granadeiro, V., Duarte, J. P., Correia, J. R., and Leal, V. M. (2013).
impact and similarly should be taken account. However, it is im- “Building envelope shape design in early stages of the design process:
portant to note that a high number of objectives and decision var- Integrating architectural design systems and energy simulation.” Autom.
iables exponentially expands the search space, and consequently Constr., 32, 196–209.
increases the convergence time. Honiden, T. (2004). “Tree structure modeling and genetic algorithm-based
approach to unequal-area facility layout problem.” Ind. Eng. Manage.
Syst., 3(2), 123–128.
Acknowledgments Inoue, M., Unehara, M., Yamada, K., Hiramoto, M., and Takagi, H. (2014).
“Evaluation of hybrid optimization with EMO and IEC for architectural
This work was supported by Middle East Technical University floor planning.” Proc., Soft Computing and Intelligent Systems (SCIS),
(METU) Bilimsel Araştırma Projesi-1 (Scientific Research Project) 2014 Joint 7th Int. Conf. on and Advanced Intelligent Systems (ISIS),
15th Int. Symp. on Advanced Intelligent Systems (ISIS), IEEE,
(BAP-1) Grant BAP-08-11-2013-070. The authors would like to
New York, 54–61.
express their sincere gratitude to T. Kaya and H. Aydin for tool
Islier, A. (1998). “A genetic algorithm approach for multiple criteria facility
implementation and Ekodenge A.S. for their contributions to the layout design.” Int. J. Prod. Res., 36(6), 1549–1569.
case study. Jones, J. C. (1992). Design methods, Wiley, New York.
Kämpf, J. H., and Robinson, D. (2010). “Optimisation of building form for
solar energy utilisation using constrained evolutionary algorithms.”
References Energy Build., 42(6), 807–814.
Kochhar, J. S., Foster, B. T., and Heragu, S. S. (1998). “HOPE: A genetic
Aiello, G., La Scalia, G., and Enea, M. (2012). “A multi objective genetic algorithm for the unequal area facility layout problem.” Comput. Oper.
algorithm for the facility layout problem based upon slicing structure Res., 25(7), 583–594.
encoding.” Expert Syst. Appl., 39(12), 10352–10358.
Koenig, R., and Schneider, S. (2012). “Hierarchical structuring of layout
Al-Hakim, L. (2000). “On solving facility layout problems using genetic
problems in an interactive evolutionary layout system.” Artif. Intell.
algorithms.” Int. J. Prod. Res., 38(11), 2573–2582.
Eng. Des., Anal. Manuf., 26(02), 129–142.
ASHRAE (American Society of Heating, Refrigerating and Air-
Lee, K.-Y., Roh, M.-I., and Jeong, H.-S. (2005). “An improved genetic
Conditioning Engineers). (2009). Climate design data ASHRAE
algorithm for multi-floor facility layout problems having inner structure
handbook, Atlanta.
walls and passages.” Comput. Oper. Res., 32(4), 879–899.
Azadivar, F., and Wang, J. (2000). “Facility layout optimization using sim-
ulation and genetic algorithms.” Int. J. Prod. Res., 38(17), 4369–4383. Lin, S.-H. E., and Gerber, D. J. (2014). “Designing-in performance: A
Balakrishnan, J., and Cheng, C. H. (2000). “Genetic search and the dy- framework for evolutionary energy performance feedback in early stage
namic layout problem.” Comput. Oper. Res., 27(6), 587–593. design.” Autom. Constr., 38, 59–73.
Banerjee, A., Quiroz, J. C., and Louis, S. J. (2008). “A model of creative Mak, K., Wong, Y., and Chan, F. (1998). “A genetic algorithm for facility
design using collaborative interactive genetic algorithms.” Design Com- layout problems.” Comput. Int. Manuf. Syst., 11(1), 113–127.
puting and Cognition’08: Proc., Third Int. Conf. on Design Computing Ouarghi, R., and Krarti, M. (2006). “Building shape optimization using
and Cognition, Springer, Atlanta, 397–416. neural network and genetic algorithm approach.” ASHRAE Trans.,
Baušys, R., and Pankrašovaite, I. (2005). “Optimization of architectural 112(1), 484–491.
layout by the improved genetic algorithm.” J. Civ. Eng. Manage., Rajasekharan, M., Peters, B. A., and Yang, T. (1998). “A genetic algorithm
11(1), 13–21. for facility layout design in flexible manufacturing systems.” Int. J.
Bleil de Souza, C., and Tucker, S. (2015). “Thermal simulation software Prod. Res., 36(1), 95–110.
outputs: A framework to produce meaningful information for design Rodrigues, E., Gaspar, A. R., and Gomes, Á. (2013). “An approach to the
decision-making.” J. Build. Perform. Simul., 8(2), 57–78. multi-level space allocation problem in architecture using a hybrid
Bogenstätter, U. (2000). “Prediction and optimization of life-cycle costs in evolutionary technique.” Autom. Constr., 35, 482–498.
early design.” Build. Res. Inf., 28(5-6), 376–386. Shayan, E., and Chittilappilly, A. (2004). “Genetic algorithm for facilities
Buscher, U., Mayer, B., and Ehrig, T. (2014). “A genetic algorithm for the layout problems based on slicing tree structure.” Int. J. Prod. Res.,
unequal area facility layout problem.” Operations Research Proc. 2012: 42(19), 4055–4067.

© ASCE 04017025-17 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., -1--1

 Tam, K. (1998). “Solving facility layout problems with geometric Wang, W., Zmeureanu, R., and Rivard, H. (2005b). “Applying multi-
constraints using parallel genetic algorithms: Experimentation and find- objective genetic algorithms in green building design optimization.”
ings.” Int. J. Prod. Res., 36(12), 3253–3272. Build. Environ., 40(11), 1512–1525.
Tuhus-Dubrow, D., and Krarti, M. (2010). “Genetic-algorithm based ap- Weaver, E., Long, N., Fleming, K., Schott, M., Benne, K., and Hale, E.
proach to optimize building envelope design for residential buildings.” (2012). “Rapid application development with Openstudio.” NREL/
Build. Environ., 45(7), 1574–1581. CP-5500-54998, National Renewable Energy Laboratory, Golden, CO.
U.S. DOE (U.S. Department of Energy). (2010). “6.0 input/output refer- Wu, Y., and Appleton, E. (2002). “The optimisation of block layout and
ence: The encyclopedic reference to EnergyPlus input and output.” aisle structure by a genetic algorithm.” Comput. Ind. Eng., 41(4),
U.S. Dept. of Energy, Washington, DC. 371–387.
Van Langen, P. H., and Brazier, F. M. (2006). “Design space exploration Yi, Y. K., and Malkawi, A. M. (2009). “Optimizing building form for en-
revisited.” AIE EDAM: Artif. Intell. Eng. Des. Anal. Manuf., 20(2), ergy performance based on hierarchical geometry relation.” Autom.
113–119. Constr., 18(6), 825–833.
Wang, M.-J., Hu, M. H., and Ku, M.-Y. (2005a). “A solution to the unequal Zitzler, E., and Thiele, L. (1999). “Multiobjective evolutionary algorithms:
area facilities layout problem by genetic algorithm.” Comput. Ind., A comparative case study and the strength Pareto approach.” IEEE
56(2), 207–220. Trans. Evol. Comput., 3(4), 257–271.
Downloaded from ascelibrary.org by ipek gursel dino on 03/15/17. Copyright ASCE. For personal use only; all rights reserved.

© ASCE 04017025-18 J. Comput. Civ. Eng.

View publication stats J. Comput. Civ. Eng., -1--1