You are on page 1of 25

SIMULATION http://sim.sagepub.

com/

Designing in complexity: Simulation, integration, and multidisciplinary design optimization for architecture
David Jason Gerber and Shih-Hsin Eve Lin
SIMULATION published online 9 April 2013
DOI: 10.1177/0037549713482027

The online version of this article can be found at:


http://sim.sagepub.com/content/early/2013/04/09/0037549713482027

Published by:

http://www.sagepublications.com

On behalf of:

Society for Modeling and Simulation International (SCS)

Additional services and information for SIMULATION can be found at:

Email Alerts: http://sim.sagepub.com/cgi/alerts

Subscriptions: http://sim.sagepub.com/subscriptions

Reprints: http://www.sagepub.com/journalsReprints.nav

Permissions: http://www.sagepub.com/journalsPermissions.nav

>> OnlineFirst Version of Record - Apr 9, 2013

What is This?

Downloaded from sim.sagepub.com at UNIV OF SOUTHERN CALIFORNIA on September 23, 2013


Simulation

Simulation: Transactions of the Society for


Modeling and Simulation International
0(0) 1–24
Designing in complexity: Simulation, Ó 2013 The Society for Modeling and
Simulation International
integration, and multidisciplinary DOI: 10.1177/0037549713482027
sim.sagepub.com
design optimization for architecture

David Jason Gerber and Shih-Hsin Eve Lin

Abstract
While the overall performance of buildings has been established to be heavily impacted by design decisions made during
the early stages of the design process, design professionals are typically unable to explore design alternatives, or their
impact on energy profiles, in a sufficient manner during this phase. The research presents a new design simulation metho-
dology based on incorporating a prototype tool (H.D.S. Beagle) that combines parametric modeling with multi-objective
optimization through an integrated platform for enabling rapid iteration and trade-off analysis across the domains of design,
energy use intensity, and finance. The research evaluates how the proposed method impacts design simulation processes,
by either enabling and/or disrupting the early stages of design decision making. This simulation technology is presented
through two major experiment sets: (1) a series of hypothetical cases emulating the architecture, engineering, and con-
struction (AEC) design modeling and simulation process using our integrated simulation framework and technology; and
(2) a pedagogically based experiment used for establishing benchmarks. Through these experiment data sets, both quantita-
tive and qualitative data are collected, including human designer and computational analysis speeds, quantity of generated
design alternatives, and quality of resulting solution space as defined by the evaluation metric of this research. The affor-
dances for incorporation of real world design complexity into our computational design prototype and simulation metho-
dology are discussed through both the enabling and the disruptive impact on the early stages of the design process.

Keywords
Parametric design, generative design, conceptual energy analysis, performance-based design, genetic algorithm, multidis-
ciplinary design optimization, decision support, design space exploration, domain integration, design process evaluation,
multi-objective optimization

1. Introduction and research objectives analysis is performed on a very limited set of design alter-
natives rather than to support early stage design decisions
With growing global concerns for sustainability, the where a broader range of possible more optimal solutions
anticipated performance of buildings is rapidly becoming may exist.8,9 Furthermore, design decisions are not typi-
a subject of intense interest and research as the built envi- cally made to satisfy a single objective, but are rather
ronment accounts for 48.7% of energy consumption in the made to seek out a ‘‘best fit’’ compromise between com-
United States.1,2 While it has been established that the peting objectives from the various expert domains.
overall performance of buildings is greatly impacted by Research precedents have demonstrated the potential of
design decisions made during the early stages of the adopting multidisciplinary design optimization (MDO)
design,3 design professionals are often unable to ade- methods to provide a performance feedback loop to better
quately explore design alternatives and their impact on
energy consumption upfront.4 Issues, such as tool intero-
perability among different expert domains, intensive anal- School of Architecture, University of Southern California, Los Angeles,
ysis time requirements, and limitations of design cognition CA, USA
and complexity can be considered as contributing factors
Corresponding author:
to design uncertainty, poor performing designs, and design David Jason Gerber, School of Architecture, University of Southern
cycle latency.5–7 Consequently, performance assessments California, 316 Watt Hall, Los Angeles, CA 90089-0291, USA.
are typically made after the initial design phase, where the Email: dgerber@usc.edu
2 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

support early design stage decision making.10,11 However, performance simulation and feedback, evolutionary or
precedents exploring MDO have typically employed sim- genetic algorithms for design, and multi-objective optimi-
plified geometry10,11 while precedents involving more zation (MOO) and MDO.
complex geometry have limited themselves to single
domain optimization.12 The few precedents which have
2.1. Parametric design and design exploration
explored both MDO and complex geometry have not
included the energy performance domain as a domain of Essential to this research is the understanding that increas-
interest.13 Where the energy performance domain has been ing design iteration can improve the possibility of identify-
included for optimization, the relationship between the ing higher performing designs through the exploration of
design form and the energy performance has been largely both geometric and non-geometric variables according to
excluded. Instead, the optimization process has focused on established design objectives and constraints.20,21 It has
mechanical systems14 and simple geometric modifications been previously established that parametric modeling can
such as window sizing and placement.15,16 Furthermore, facilitate rapid changes of geometric and non-geometric
the application of these precedents’ subject of interest to variables according to a pre-rationalized design
the overall design process remains largely unexplored. logic.20,22,23 Therefore, a primary building block of the
Considering the impact of early design decisions on research is the utilization of parametric design and model-
the overall building performance and the interest in under- ing for the purpose of providing rapid design iteration.24
standing the applicability of MDO in assisting design deci- The use of parametric design is becoming increasingly
sion making, a new early stage design and simulation common due to the acknowledged advantages and inherent
methodology was in need of development. As a response, ability to quickly generate design alternatives through the
we proposed and developed an integrated simulation parameterization of geometry.24,25 Once a parametric
framework which uses our proprietary prototype tool model is formed, design alternatives can be rapidly gener-
(H.D.S. Beagle) enabling the coupling of parametric design ated through parameter value manipulations. Also, the cre-
with multi-object optimization to be applied to the early ation of a parametric model requires rigorous definition of
stages of the design process.17–19 The objective of this the relationships between geometric objects which enables
paper seeks to explore the applicability of this proposed associative and relational design, i.e. correlative, integral,
framework for the architecture, engineering, and construc- and explorative design process.26 Furthermore, once
tion (AEC) design process through a series of design sce- design intent has been defined through parametric relation-
narios, hypothetical and pedagogical, as a means of ships, the automation of the exploration process becomes
observing the impact of this integrated simulation approach possible.27 The automation of the design exploration pro-
on the early stage design process. During this exploration cess not only can shorten design exploration cycle time
the impact factors of the overall simulation process were (i.e. latency) but also provides the opportunity for relation-
divided into two categories for further analysis: computa- ally incorporating performance criteria as part of the pro-
tion and automation affordances and human design process cess. This provides an opportunity to integrate more
affordances. criteria including performance objectives and constraints,
While there is a definite correlation between these two and thereby more extensive feedback by which to evaluate
categories,11 the focus of this paper will be on comparing and identify potentially optimal design alternatives. Aish
the qualitative and quantitative measurements available and Woodbury’s work28 demonstrates how the automation
through the technology and thereby exploring the design of parametric design can significantly reduce the time
required for change and reuse. Another precedent, the gen-
computational affordances to both enable and disrupt the
erative performance oriented design model (GenPOD), uti-
complex human design process and the complexity of real
lizes the parametric form generation process to generate
world design and trades offs. Once the baseline of the
various performance envelopes and demonstrates promis-
technological impact has been established, the ability to
ing support for design decision making.29 Rolvink et al.,30
identify the impact factors on the human designer provides
in a practice setting, also demonstrated the effectiveness
the research with the means of observing the affordances,
of using parametric modeling techniques in rapidly explor-
enabling and disruptive, of both categories of the our
ing design alternatives and associative structural perfor-
MDO framework to fundamentally ‘‘design-in complex-
mance. However, the essential component of parametric
ity’’ and close the performance feedback loop for early
modeling with regards to this research is the ability to pro-
stage design decision making.
vide ranges of potential parametric values that can be used
to generate solution populations of bounded though expan-
sive variation similar to biological genomic models. These
2. Background and review
solution populations can then be accessed through automa-
This research is built upon the precedent of four research tion as a means for identifying potential better solutions
domains: parametric design and design exploration, that meet complexly competing MOO criteria.
Gerber and Lin 3

2.2. Performance simulation and feedback advancement of technology and the increase in information
One such objective is that of estimated energy performance fidelity and availability, the process of design has become
more complex as opposed to less so. Consequently, multi-
through design and analysis simulation. With the continued
disciplinary design considerations have become more and
advancement of computational tools for building design,
more unavoidable. In order to deal with the increased com-
the notion of anticipated performance has gradually been
plexity, and the increase in competing objectives, a sys-
allowed to claim a more prominent role as a driving force
tematical problem solving technique is needed. In
behind design decisions.31 It must be noted that the term
numerous design and engineering fields, MDO methods
‘‘performance’’ for this paper is bounded by the objectives
have been explored as a potential approach to tackle and
of energy usage, program constraints, and financial esti-
manage these problems. MDO refers to methods to solve
mates. It is furthermore used in terms of design processes
design problems which have several objective functions
and metrics, i.e. the performance of the system to enable or
and incorporate a number of disciplines, the normative
disrupt predominant early stage integrated design and per-
case for design.37 As defined by Poloni and Pediroda,38
formance analysis practices.
MDO is achieved through ‘‘the art of finding the best com-
Environmental simulation is not new to the field of
promise’’. Previous building design precedents have inves-
architecture, but when and how it is incorporated into the
tigated the application of a multi-objective genetic
design workflow can be problematic.32 Conventionally
algorithm (MOGA) for finding the optimal in the trade-
adopted performance-based analysis methods have been
offs between capital expenditure, operation cost, and occu-
shown by prior studies to be ill suited in their ability to
pant thermal comfort in building design.39 Flager et al.10
support early stage design decisions due to time limita-
adopted a MDO method to perform a study on a simple
tions.33 In addition, there is often the issue of environmen-
classroom optimizing structural and energy performance.
tal simulation software needing to be operated by experts
Magnier et al.40 used a MOO algorithm to optimize the
due to the typically specialized nature of these tools.34 As
energy consumption and thermal comfort of a residential
a result, the designer is left with the need for coordination
building. The ‘‘CATBOT’’ project utilized MDO to link
between the input of multiple experts and therefore an
complex geometry to structural analysis.41 These prece-
increase in design cycle latency. Another contribution to
dents all demonstrate the potential abilities of MDO to
the issue of design latency is the lack of integration
assist in identifying higher performance solution sets
between tools used for different environmental simula-
among multiple competing criteria. At present, however,
tions.35 Restricted data exchanges and multiple interfaces
most of the precedent work have either had singular
required for multidisciplinary design problems are ongoing domain emphasis on structural performance, detailed
issues still in need of resolution.36 Augenbroe states, ‘‘the mechanical systems, or have been applied to simplified
role of simulation tools in the design and engineering of geometric settings. As well, the application of preliminary
buildings has been firmly established over the last two energy performance feedback to support complex geome-
decades.’’ Critical for our research, Augenbroe continues try has not been fully understood and therefore developed.
to credit energy simulation ‘‘with speeding up the design Our review of the literature and practice in fact enumerates
process, increasing efficiency, and enabling the compari- this significant gap given the contemporary need for better
son of a broader range of design variants, leading to more incorporation of energy related factors into design simula-
optimal designs’’ and in summary providing better under- tion of complex geometry or in other words contemporary
standing of trade-offs in the multidisciplinary design pro- designs.
cess.6 However, despite the established use of simulation
software, there are still issues persisting regarding
increased design cycle latency and lack of integration 2.4. Genetic algorithm for design
across other domains and their analysis environments. Genetic algorithms (GAs) are based on the evolution by
This research proceeds to explore our previously estab- natural selection model formulated by Charles Darwin42
lished methodology18 to observe if these issues were and have recently drawn the attention of design computing
addressed and to determine if the tested methodology was researchers interested in optimizing multiple performance
an efficient means of integrating estimated energy perfor- criteria.43–45 Since the pursuit of design alternatives within
mance and financial performance during the early stages the design process is often nonlinear, stochastic approaches
of design as part of the MDO process. such as GAs are typically considered suitable.46 The GA
was first introduced by John Holland47 in the 1970s as a
heuristic search method with the capacity to have broad
2.3. MDO applicability, ease of use, and the capacity to explore
Design is a multidisciplinary process by nature that is akin numerous variables from a global perspective.48 Frazer43
to evolution, a balancing act between competing objectives proceeded to apply this theory to the field of building
all vying for the greatest influence. Furthermore, with the design where the ability of GAs to generate a large
4 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

solution pool addressing multiple variables seemed suit- (SPC). To this end, conceptual geometric designs were
able. Gero49 is also credited for the application of GAs in integrated through the prototype with energy use estimates
design research for his introduction of utilizing environ- and preliminary economic appraisals into an automated
mental performance for multi-objective design problems optimization routine by means of parameterization, plat-
and Pareto optimization for enabling design decision form integration, and MOO. Through the enabling of auto-
making. In addition, GAs have been researched in the matically generated design alternatives in accordance to
application of optimizing overall energy performance, by user defined parameter ranges, combined with concurrently
Tuhus-Dubrow and Krarti,50 who developed a simulation- generated energy analysis results of each design alterna-
optimization tool that couples a GA with building energy tive, H.D.S. Beagle allows for an automated calculation of
simulation in order to optimize building shapes and envel- three objective functions that are then utilized by a custo-
opes. In another study a multi-objective GA method was mized GA-based MOO algorithm to search, rank, select,
applied to focus on economical optimization by minimiz- and breed the solution space as defined by the GA off-
ing energy consumption.51 All of these studies support the spring scoring and parameter settings.
need for a continued development of systematic optimiza-
tion algorithms that can handle large solution sets with
3.1.1. Tool description and platform process integration. To
multiple competing objectives in order to generate a more
achieve the desired research objective, the software plat-
optimum solution set. This research builds upon the use
forms AutodeskÒ RevitÒ and AutodeskÒ Green Building
and integration of a GA for design optimization with the
StudioÒ (GBS) were selected based on existing interoper-
ability to generate a set of better-fit solutions. The intent is
ability capabilities to be combined with MicrosoftÒ
not to identify an optimal solution, which is often not pos-
ExcelÒ into an automation and optimization routine. The
sible, per se, but to provide an environment in which
overall system architecture is illustrated in Figure 1.
informed trade-off decisions can be made more rapidly
AutodeskÒ RevitÒ is a building information modeling
and with more certainty by the computational and human
platform with some parametric capabilities enabling
multidisciplinary design team.
designers to define their geometry and provide a series of
parameters that impact the development of varying geo-
3. Research method metric configurations. This platform also serves as an
insertion point for the energy settings necessary for a con-
In order to effectively evaluate the MDO framework estab-
ceptual energy analysis through GBS. This approach
lished by precedent research and by our own prior
avoids the hurdle of geometry and parameter translation
research,18 and the impact this has on the simulated perfor-
between the design and energy simulation domains. GBS
mance feedback loop during the early stages of design,
is a web-based energy analysis service that serves as the
evaluation metrics needed to be established. In response,
energy simulation engine for H.D.S. Beagle. The Beagle
after the development phase of the prototype tool, the
also uses MicrosoftÒ ExcelÒ 2010 to provide, not only a
research proceeded to determine measurement metrics to
means of containing the financial parameters and formulae,
be used as a standardized guideline to provide a means of
but also as a user interface (UI) proxy in which designers
evaluating and analyzing collected data. Our methodology
can set up design parameter variation ranges, constraints,
includes the following: precedent review, prototype specifi-
spatial program parameters, and the SPC formula. In order
cation and building, implementation and experimentation,
to integrate these three expert domains and platforms, the
re-running and re-designing of experiments in a controlled
Beagle was developed in C# as a plugin for AutodeskÒ
and more measurable fashion, metric establishment, data
RevitÒ by using the Revit API, GBS SDK, and Excel API,
collection and comparison, and finally review of next steps.
as shown in Figure 2.
In this section, the prototype tool is summarized, followed
by the basis on which the evaluation metrics are built.
Finally, the data sets from both the hypothetical cases and 3.1.2. Parameterization and problem formulation. The
the pedagogical benchmark cases are described. research requires parametric models to be designed to gen-
erate a large pool, i.e. a solution space of alternatives. In
order to automatically generate a solution space, it is nec-
3.1. Summary of the prototype tool (H.D.S. Beagle) essary to first formally define the design problem into a
In order to explore the research objectives, a MDO design series of design objectives, variables, and constraints.
framework utilizing prototype tool (entitled H.D.S. Beagle) These definitions are then used to generate an associative
was developed through prior research.17–19 H.D.S. Beagle parametric design model, which implicitly describes a
associates parametric modeling, design domain integration, bounded and a topologically fixed solution space. In order
and a GA-based MOO algorithm with specific focus on to create a flexible yet defined design workflow, there are
energy use intensity (EUI), financial performance net pres- a total of four parametric categories used within the
ent value (NPV) and spatial programing compliance research framework: design geometry parameters, energy
Gerber and Lin 5

Figure 1. The system architecture overview of H.D.S. Beagle 1.0.

Figure 2. H.D.S. Beagle automation loop developed through C# as a plugin for Revit, using Revit API, Excel API, and GBS SDK.
6 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

setting parameters, spatial program compliance para- mutation, the population is able to generate more diversi-
meters, and financial pro forma parameters. Among these fied offspring. This three-step process is then repeated,
four parameter groupings only design geometry parameters with each cycle representing one step of the evolution, i.e.
and energy setting parameters can be considered as geno- a new generation. The process is cyclical and continuous
type parameters to be used as ‘‘genes’’ later while spatial until the algorithm converges or other stopping criteria are
program parameters and financial pro forma parameters reached. Currently, the population method, crossover ratio,
are used as a means of calculating the phenotype objective mutation rate, selection size, and maximum iteration num-
functions used in the evaluation of offspring during the ber are set manually by the designer in the Beagle in order
optimization process. This is due to spatial program para- to accommodate for user preference and due to varying
meters and financial pro forma parameters being directly complexity, coupling – i.e. interaction of parameters, and
dependent on design geometry parametric settings and scale – i.e. number of parameters of design problems.
energy settings, thereby possessing no driving parameters The ‘‘fitness criteria’’ of the offspring is determined by
of their own. Design geometry parameters are treated as H.D.S. Beagle based on compliance with defined objec-
design problem-specific parameters to be specified by the tives, i.e. targets and constraints, from the three integrated
designer. This requires designers to provide a basic geo- design domains: design compliance, energy performance,
metrically variable model in AutodeskÒ RevitÒ that is pur- and financial performance. The categories are then used to
posefully built to enable configuration generation. Energy rank design solutions and identify design solutions with
setting parameters are determined by the designer and are the potential for optimization or trade-off analysis in all
available through the AutodeskÒ RevitÒ’s conceptual three categories. The definition of the objective functions
energy analysis built in settings. Spatial program para- used in determining ranking within these three categories
meters are the parameters used for calculating the SPC is deliberately bracketed within H.D.S. Beagle in order to
score, e.g. the amount of floor area per program type, that evaluate the effectiveness of this simulation and evaluation
are also set by the designer. Financial pro forma para- methodology. For example, the aesthetic of the design is
meters are based on a financial calculation model in order subjective in nature and therefore does not possess the
to determine the NPV of each design offspring. ability to be ranked quantitatively in a convenient, accu-
rate, and effective fashion. In response, H.D.S. Beagle’s
3.1.3. GA-based MOO. In H.D.S. Beagle, genes are equal determination of ‘‘Spatial Programming Compliance
to modifiable parameters as bounded continuous variables, (SPC)’’ was constrained to the use of meeting specified
and chromosomes correspond to individuals that are com- design program requirements. In order to determine energy
posed of these genes. The customized GA then optimizes performance, the EUI provided by GBS was chosen as the
an initialized set of individuals (population) using three objective function representative of the energy perfor-
main steps: (1) evaluation, (2) selection, and (3) popula- mance category. The objective is to identify and explore
tion. These solutions are then evaluated based on a series the potential design solution with the lowest calculated
of ‘‘fitness criteria’’. The ‘‘fitness criteria’’ are defined in EUI. The financial performance’s primary goal was to
H.D.S. Beagle by the objective functions of the design reflect the expected initial cost of different geometry con-
solution defined below. figurations, potential returns, and maintenance costs. To
provide this information, the program extracts relevant
Sobj = Max. SPCS
information from proposed geometry and translates it
Eobj = Min. EUI
into an expected financial cost generated by provided
Fobj = Max. NPV
material costs. For the purpose of the initial experiments
where a simplified version of the NPV formula is utilized to
provide broad estimates regarding the cost of construc-
Sobj = SPC Objective Function
tion, operation costs, and expected revenue values.
Eobj = Energy Performance Objective Function
However, when more detailed information is integrated,
Fobj = Financial Performance Objective Function
a more accurate estimate of costs and revenues can be
SPCS = Spatial Programming Compliance Score
generated. In the current version of the H.D.S. Beagle,
EUI = Energy Use Intensity
there are two mechanisms that will stop the GA; the first
NPV = Net Present Value
is defined by the user as the maximum iteration number
The closer to fulfilling all ‘‘fitness criteria’’ the more when initializing the GA, and the second is defined when
‘‘fit’’ a design solution is considered. The more ‘‘fit’’ a the GA reaches the convergent criteria. The convergence
solution is, the higher the probability that the solution will criteria are defined when there are three generations that
be selected to ‘‘survive’’ in the next generation, using a have the same optimal result, i.e. there no longer is a
tournament selection method.52 The next generation is then quantifiable improvement or difference. At this point,
populated by these ‘‘survivors’’ through a recombination the GA will determine the design has reached the opti-
of genes and mutation. Through this recombination and mal solution.
Gerber and Lin 7

Figure 3. The six step process for integrating design and energy simulation independent of platform or tool use.

3.2. Process evaluation specifically for the purpose of evaluating these processes
To measure the effectiveness of the MDO framework in and their enabling or disruptive properties on the early
the early stage of design process, an evaluation metric is design process.
needed in order to isolate the means by which our simula- In order to consider both the quantitative and qualita-
tion framework can be measured against existing simula- tive values in application to the integration of design and
tion frameworks. A broad spectrum of disciplines has energy simulation process, while maintaining consistent
taken an interest in dissecting, analyzing, and measuring results across different project activities, this study nar-
the design process in order to identify possible means of rowed the scope of design activities of interest to six steps,
improvement. The work of Eastman,53 Akin,54 Cross as illustrated in Figure 3.
et al.55 and Gero et al.56 are the early representative efforts The identified six steps for energy simulation are as
in the architectural design field that attempted to measure, follows:
describe, and analyze design protocol and activities.
Following this work, Kalay57 and Stoyell et al.58 started to 1. Generate design configuration
address the performance-based design aspect of design 2. Transfer design model to energy simulation model
process by measuring the impact of design activities on 3. Modify energy model and apply energy related
the building lifecycle performance. There are also research attributes
that evaluate robustness and the flexibility of the early 4. Run analysis
stage design exploration process.59 Recently, the work of 5. Evaluate results
Clevenger et al.60,61 emphasized the need of providing a 6. Execute design decisions based on available
metrics by which to assess design guidance. However, as feedback
summarized by Clevenger et al.,
‘‘In general, limited real-world data exist to evaluate These steps can either be manually implemented or
the exploration performance achieved by actual designers. implemented automatically depending on the user’s
This is due, primarily, to the fact that parallel or redundant selected tools and platforms.
explorations are not performed across strategies in the real These six steps are applicable even within the context
world due to limited project resources.’’61 of platforms which combine the design geometry and
In addition, the majority of these studies examine the energy simulation domains. When such a platform is uti-
design process from a macro scope approach, attempting lized, steps 2 and 3 can be considered as being completed
to link individual activities with overall design results. through an automated process rather than a manual one. In
However, the subjective nature of the design process is order to evaluate the effectiveness of the six step process,
highly reliant on individual preferences and thereby resists various factors must be taken into consideration.
these top-down measurements as a means of providing Referencing Atman et al.’s categorization method,62 this
clarity regarding the process evaluation. The subject of research proceeded to categorize the types of measure-
interest for this research is on the integration processes ments to be collected into four categories: design problem,
between geometric design and energy performance feed- process, product, and actor domain and experience, as
back for the purposes of early design decision making. listed in the Table 1. The purpose was to provide a basis
Therefore this research proceeded to establish metrics by which to observe the impact of both technical and
8 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

Table 1. Evaluation metrics used in our study.

Categories/Measures Record Unit/Method Hypothetical Cases Pedagogical Benchmark Cases

Design Problem Measures


Project complexity
Project size Sqft O O
Space type number Number O O
Design complexity
Energy model surface count Number O O
Explored parameter numbers Number/descriptive O O
Process Measures
Speed
Time spent to create design geometry Minutes O O
Time spent to run energy analysis Minutes O O
Time spent to calculate three objectives Minutes N/A O
Feedback method measures
Performance feedback time per result Hour O O
Manual/Automated M/A O O
Simulation process observation
6 step implementation Diagram, Descriptive O O
Exploration process measures
Decision making patterns Diagram O O
GA Settings Number O N/A
Product Measures
Feedback quantity measures
Feedback number per day (8 hours) Number O O
Feedback quality measures
Feedback solution quality EUI, NPV, SPC O O
Output information for each result Descriptive O O
User Experience
Parametric model experience Descriptive O O
H.D.S. Beagle experience Descriptive O N/A

human factors together so as to determine the effective- scores to become available for feedback purposes, i.e.
ness or potential value of our simulation framework. One enabling a design decision or choice and whether the
issue to note is that the developed evaluation metric was method of obtaining the objective scores was either manual
limited to the available measurable items collected or automated. Simulation process observation describes
through the exploration of our hypothetical design scenar- the implementation method used when completing the pre-
ios and pedagogical benchmark case. viously described six step energy simulation process.
Measurements falling into the design problem category Exploration process provides diagrams illustrating the
are values collected regarding the physical aspects of the decision making patterns used by the hypothetical case
design and are further divided into two subcategories: proj- studies and the pedagogical benchmark cases. Product
ect complexity and design complexity. Project complexity measures focus on evaluating the resulting design alterna-
refers to the project size as measured in square feet and the tives of each process of interests. These include the num-
number of types of program spaces, such as parking, com- ber of design alternatives generated during an eight hour
mercial or residential, that are included within the design work period and the solution quality as defined by EUI,
problem. Design complexity refers to the amount of sur- NPV, and SPC. The user experience section documents the
faces required to be included in the energy model along users’ background experience in utilizing parametric mod-
with the number of available parameters as provided by eling and the previous experience using H.D.S. Beagle.
the design problem. Process based measurements are
divided into four subcategories: speed, feedback, simula-
tion process observation, and exploration process. Speed 3.3. Data collection
measures times experienced during the specified activities In order to evaluate our simulation framework, the
of creating the design geometry, running the energy analy- research conducted two types of experiments: hypothetical
sis, and calculating the three objective scores. The feed- and pedagogical case scenarios. Hypothetical cases are
back category records the time taken for all three objective composed specifically for use by H.D.S. Beagle with
Gerber and Lin 9

Figure 4. Implemented prototype, i.e. computational, and pedagogical, i.e. manual, simulation process outlines.

varying degrees of geometric complexity. These serve as a in Figure 4 is the adjusted simulation process used for the
means to capture the simulation process implemented pedagogical benchmark experiments. The significant
through the design methodology utilizing the Beagle. A simulation process difference is predicated on the fact that
series of pedagogical experiments are then run in a class- the pedagogical experiments were run in a classroom set-
room environment to use as a means of establishing ting where students were not granted access in real time
benchmark references regarding the following: run time, and in parallel to the automation characteristic of the
geometric complexity, feedback analysis, and quality of Beagle. The last three steps are performed manually as
the solution space. The benchmark cases are also used to opposed to automated through the GA-based MOO pro-
observe how our design methodology enabled or disrupted cess provided though the Beagle.
the conventional manual design process simulated in the The following outlines the experimental parameters,
pedagogical case. data of interest, simulation and exploration process, col-
lected data, and observations for the two sets of experi-
ments contained in this research.
4. Case studies and data analysis
As previously discussed in Figure 3, there is a six step
design simulation process for integrating the design geo- 4.1. Hypothetical cases
metry and energy simulation feedback. When applied to 4.1.1. Experiment description. In order to explore the proto-
our MDO methodology, the resulting simulation process type’s functionality, limitations, and applicability to real
for the prototype is illustrated in Figure 4. Also illustrated world scenarios a series of hypothetical design problems
10 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

are generated then run in accordance to each problem’s Step 3 Modify energy model. The step is bypassed since
specifications. These 12 case studies are developed by the the model transfer to GBS is automated by Revit, which is
authors to be implemented through the use of H.D.S. unavailable through most non-integrated platforms.
Beagle. These experiments are designed to test computa- Step 4 Run analysis. This step has two stages. The first
tion and automation affordances and to emulate the early stage is to send the automatically translated energy model
stage simulation process as used by our framework in of the design alternative to GBS for analysis and obtain
order to observe the enabling and disruptive properties of the results. This part is executed through Revit by the
the overall process. Beagle after a design alternative is generated. Once the
energy analysis results are available the Beagle then pro-
4.1.2. Simulation process. The simulation process of H.D.S. ceeds with extracting the relevant information from the
Beagle as applied to the previously defined six step simula- results and the design model to automatically calculate the
tion is illustrated in Figure 4. The resulting detailed work- SPC score and the financial performance score of the
flow for the hypothetical cases is described below. design alternative. These scores are then put with the EUI
score provided by GBS for the next evaluation step.
Step 1 Generate design. For this process there are two Step 5 Evaluate results. Performance of design alterna-
subcategories: the generation of the initial design and the tives within each generation are automatically evaluated,
generation of the design alternatives. The initial design is ranked, and scored before candidates for the following gen-
developed through the preparation of the initial executable eration are selected by the GA. Figure 7 illustrates an exam-
design and the constraints file according to the designer ple of the Beagle’s genetically driven evaluation process
driven project requirements. At this point, each hypotheti- where the initial user defined population of 10 was popu-
cal case is provided a specific site, weather and climate lated followed by three subsequent generations bred based
information, and overall program objectives to simulate on the performance of the previous generation. Once the
real world design problems. This step is also where para- user requested number of generations has been reached then
metric ranges are set. These ranges are set once per GA the results are provided to the user for manual evaluation.
run. The generation of design alternatives is then produced At this point, the GA run is considered to be completed.
through the automated process of exploring combinations Step 6 Execute decision. Once the results have been
and mutation percentages of parametric values within their generated there are two ways to proceed: (1) a designer
set ranges and subsequent performance results calculated manually implements changes in the initial design of the
by the GA. A total of 12 distinct design scenarios are pre- executable design file based on the acquired simulated
sented in this paper. They range from single program type results, which would also suggest another optimization
to up to four mix–use program combinations. These sce- run; or (2) a design alternative is selected based on the
narios were also generated using varying levels of geo- multi-objective trade-off analysis provided by the Beagle
metric complexity. A summary of the 12 scenarios is and the design proceeds to the next stage of development.
provided in Figure 5. Each case has further been defined For the purpose of these experiments, step 6 was executed
through a series of site constraints, program requirements, through identifying the higher performing design alterna-
and geometrically driving, fixed, and driven parameters, tives and their performance ranges for the SPC, EUI, and
and formal, i.e. geometric complexity. The geometric NPV scores as provided by the Beagle.0
complexity of each problem has been measured through
the surface count of the initial design’s energy model as 4.1.3. Experiment results and observations. A summary of
provided by the energy simulation software. Figure 6 pro- the collected data from the 12 hypothetical scenarios is
vides a sample example of the components and attributes provided in Table 2.
of these scenarios. The acquired quantitative measure- During the course of running the 12 hypothetical sce-
ments through the established metrics are summarized in narios, several observations were made. When utilizing the
Table 2. prototype there is a direct correlation between the run time
Step 2 Transfer model. The integrated platform enables and the surface count of the energy model. This linear rela-
the direct translation of the design geometry and related tionship is described in Figure 8. Currently, the averaged
energy settings into the energy simulation engine provided time for each surface to be calculated was approximately
by GBS. As a result, an analyzable energy model can be .0036 seconds. While the correlation between the surface
obtained directly thereby proceeding with the energy count and the run time is evident, there is no observed cor-
simulation without additional modification of geometry or relation between the time required for the energy analysis
energy related attributes. In this process, the prototype and the quantity of parameters, e.g. the scale of the design
automatically converts and sends the design alternative to problem intrinsic to each design scenario. However, there
the GBS server to request and obtain a conceptual energy may be a correlation between the quantity of the para-
analysis. meters and the overall time necessary in reaching the
Table 2. Summary of the hypothetical cases measures.

Categories/Measures Scenario No.

1 2 3 4 5 6 7 8 9 10 11 12
Gerber and Lin

Design Problem Measures


Project complexity
Project size (sq ft) 16768 84680 167680 16500 30310 51000 3000 86000
Space type no. 4 4 4 4 1 2 4 1 1 3 1 1
Design complexity
Energy model surface counta 794 1587 1042 824 7517 2451 270 1060 28 2086 197 338
Explored parameter no. 6/61 12/3 16/3 23/3 7/27 8/0 10/3 13/3 0/12 7/4 9/1 6/21
(Design/Energy)
Process Measures
Speed
Time spent to run energy 1.46 3.25 2.95 2.96 28.64 8.36 1.01 2.90 0.53 4.15 0.91 1.70
analysis (minutes)b
Feedback method measures
Performance feedback 2 7 8 8 41 40 2 3 1 6 1 4
time per result (minutes)c
GA setting
Initial population 20 10 10 10 20 10 10 10 40 10 10 20
Crossover ratio 0.5 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6 0.6
Mutation ratio 0 0 0 0 0 0.006 0 0 0.006 0 0 0.006
Population size 40 20 20 20 40 10 40 20 40 20 20 40
Selection size 30 20 20 20 30 10 30 20 20 20 20 10
Maximum iteration 20 0 0 20 1 10 40 20 20 6 5 10
Product Measures
Feedback quantity
Feedback number per day 240 34 30 30 5 6 120 80 240 40 240 60
(8 hours)
Feedback quality: (Initial/Solutions’ range)
NPV (Million Dollars) 528155~754 7116~71 92N/A 13274~525 538142~555 (-94)(-516)~84 73876~741 565113~769 (-73)(-74)~(-71) (-41)(-40)~834 (-3)(-4)~(-2) 34(-57)~178
EUI (kBtu/sq ft/yr) 5545~68 5651~77 63N/A 6243~83 6555~88 5752-67 4842~88 6149~79 5653~104 17356~233 6451~99 5447~99
SPC 7524~94 60~50 3N/A 55~88 8838~99 54(-76)~71 3131~95 833~100 100N/A 10(-404)~88 9946~100 9948~99
a
The surface count is according to the energy model of the initial design geometry. During the GA process, varying design options will have varying surface counts.
b
These time measurements were according to generating the initial masses’ energy models and include the time required to both transfer to and receive results from Green Building Studio through the
Internet.
c
These recorded times include time needed by H.D.S. Beagle to update new design options, calculate three objective functions evaluating each design option, and proceed with tournament selection.
11
12 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

Figure 5. Summary of the 12 hypothetical design scenarios including program complexity, parameter scale and coupling, and
geometry.

mathematical converging optimal design solution pool. the run time; we believe this is also in need of further
This possible correlation is in need of further explora- observation.
tion. Furthermore, according to the present data, there There are also some technological limitations observed
was no observed correlation between the project size and during this set of experiments. First, the complexity of the
Gerber and Lin 13

Figure 6. Example of hypothetical design scenario composed of site attributes, program requirements, and the three parametric
categories, driving, driven and fixed.

Figure 7. Illustrates the GA driven evaluation by H.D.S. Beagle of hypothetical scenario 10 and a subset of the resulting solution
pool.

project is limited by the GBS’s energy simulation engine. tool, this process was done automatically. This means that
The selected energy simulation engine can currently only when a designer designated a geometric element as an
analyze 8192 exterior surfaces, 8192 interior surfaces, exterior wall in the design platform, the exterior walls’
8192 underground surfaces, 1024 shade surfaces, 8192 attributes were preserved and transferred into the energy
openings, and 4096 spaces.63 As a result, if the energy model without further manipulation. However, when a
model of the designed geometry exceeded any of these design model possesses higher degrees of curvature, such
limits the automation loop is disrupted. Also, there was an as scenarios 2 and 3, there is the possibility of a mistran-
issue experienced with regards to the process of transfer- slation. For example, a curved roof surface might be
ring a design model to an energy model. In the prototype recognized as an exterior wall surface during this transfer
14 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

Figure 8. Illustration of the relationship between the geometric complexity and the analysis runtime. Geometric complexity is
defined as the energy model surface count of the initial parametric model design.

process. This means that in the energy model the surface simulation programs. Also, most significantly H.D.S.
would have the attributes of the exterior wall instead of the Beagle is able to successfully provide a design solution
attributes assigned to the roof material. This mistransla- pool with an improved performance over the initial design
tion, in turn, affects the accuracy of the analyzed results for all 12 scenarios. These results have been further dis-
that are dependent on these attributes to execute the energy cussed in section 5.3.
performance calculations. The occurrence of these types of
mistranslations can be manually identified and corrected
by examining the resulting energy model. However, during 4.2. Pedagogical case
the automated GA process it is impossible to manually 4.2.1. Experiment description. In order to observe the impact
correct any identified issues in the generated energy mod- of the proposed simulation methodology on design cycle
els for each design alternative. The extent of the impact latency and performance feedback latency, a pedagogical
this type of error has on the accuracy of the generated experiment was developed to generate a series of bench-
results and how this inaccuracy affects the GA process is marks for comparison purposes. 27 students participated in
in need of further investigation, though we believe it to be this study after receiving seven weeks of Revit training.
statistically not significant as the number of surfaces this The study group was composed of 18 master of architec-
mistranslation effects is normally small and only in the ture candidates, 8 master of building science candidates,
complex geometry scenarios. Despite these issues the and 1 undergraduate student pursuing a minor in architec-
Beagle is able to successfully enable the exploration of ture. For this experiment, hypothetical design scenario 10
more complex geometry, as shown through scenarios 2, 3, was provided to the students with the following design
and 5, than typically modeled in common energy objectives and constraints:
Gerber and Lin 15

Figure 9. Initial parametric model provided to each student for the pedagogical experiment.

Site location: 8782 Sunset Boulevard West Hollywood, Students were given nine parameters by which to explore
CA 90069 potential design alternatives; five parameters drove the
Site Area: 22,500 Sq. ft. (150’ x 150’) overall building form and geometry and four parameters
Design Requirement: were built in energy parameters available through the
Retail Area 11,000 Sq. ft. Revit conceptual energy analysis platform. The provided
Office Area 20,000 Sq. ft. model and parametric ranges are illustrated in Figure 9.
Hotel Area 20,000 Sq. ft. Figure 10 provides a sample of the worksheet given to stu-
dents to record their progress.
Design Goal/Objectives: Step 2 Transfer model. Once a design alternative has
a. Maximize the SPC Score been generated students are asked to analyze the perfor-
b. Minimize EUI mance of their new design option. Students are instructed
c. Maximize NPV to generate an energy analysis through Revit using the
conceptual analysis tool that sends and retrieves the
energy analysis from GBS.
4.2.2. Simulation process. At the beginning of the experi-
ment, each student was provided an identical Revit model Step 3 Modify energy model. This step is bypassed as
with predefined parametric options to start with. Ranges the model transfer is automatically provided by Revit to
for each parameter setting were also provided. Students GBS.
were then asked to explore multiple combinations of these Step 4 Run analysis. Energy use analysis results pro-
settings in order to generate a potential design solution vided by GBS are provided automatically once the con-
with the lowest calculated EUI, the maximum financial ceptual energy analysis is requested through Revit. Once
performance (NPV), and the maximum SPC. The EUI was acquired, students then record the EUI levels in the pro-
calculated through the service provided by Revit which vided Excel worksheet along with other design attributes
automatically translated the conceptual mass into a sche- as explained in their experiment set up and illustrated
matic energy model. Financial performance and SPC were Figure 11. The input from the students is then used to
calculated through Excel files provided to the students score their design alternative according to the three previ-
which were identical to those used by H.D.S. Beagle to ously defined objectives; EUI, SPC, and NPV.
generate similar calculations. According to the previously Step 5 Evaluate results. Design attributes and the result-
described methodology, the students were asked to follow ing scores are then documented in the provided Excel
these six steps: worksheet to allow for comparison against previously gen-
erated results, as illustrated in Figure 12. This analysis is
Step 1 Generate design. In this experiment, the initial then available for students to review before proceeding
design was provided so the generation of design alterna- with any changes in the design that is of interest.
tives was based on the manual manipulation of the pro- Step 6 Execute decision. Modifications of the design
vided parameters within their set parametric ranges. are made according to the designer’s interpretation of the
16 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

Figure 10. The worksheet provided to students to record their decision making process.

Figure 11. The Excel worksheet calculator provided to students used to calculate the performance of their generated design
alternative according to the three objective functions.

generated results. During this period, students are asked to 4.2.3. Experiment results and observations. This study col-
record their progress along with the time spent during lected data from 27 students who generated a total of 118
each process stage: different design options through the provided parametric
model with a minimum of three iterations per student.
1. Updating the design option: time recorded to mod- Table 3 presents each student’s maximum iteration num-
ify geometry, level settings, and space settings. bers, time spent on the overall process, and their resulting
2. Energy analysis: time recorded to request the con- solution space range. Lastly, Figure 13 illustrates the stu-
ceptual energy analysis for the design alternative dent’s resulting data points and their exploration samples.
through Revit and to receive the analysis results For this experiment, the students were asked to opti-
from GBS. mize the design according to the same three objectives
3. Objective calculations: time recorded to manually used by the Beagle: minimize EUI, maximize NPV, and
input the necessary data into the Excel calculator maximize SPC. As summarized in Table 3, the average
in order to obtain the objective calculations exploration time recorded was three hours with approxi-
(including the time previously recorded for the mately five generated iteration numbers. There was no
energy analysis). observed correlation between the solution quality, iteration
Gerber and Lin 17

Figure 12. The worksheet provided to students to allow them to compare multiple design iterations.

Table 3. Summary of pedagogical iteration numbers, exploration times, and the explored ranges of generated solution space.

Student Iteration Exploration NPV (Million Dollars) EUI (kBtu/sq ft/yr) SPC
ID No. Time (Hour)
MIN MAX RANGE MIN MAX RANGE MIN MAX RANGE

S1 4 5 − 41.0 − 8.8 32.1 70.4 174.9 104.5 9.6 35.2 25.5


S2 4 4 − 29.9 138.8 168.7 81.0 246.8 165.8 18.1 97.9 79.9
S3 4 3.5 53.7 138.1 84.4 66.8 137.2 70.4 57.4 85.7 28.3
S4 4 1 − 32.9 60.9 93.8 79.4 173.0 93.6 16.1 89.3 73.3
S5 4 4.5 − 22.1 114.4 136.5 91.9 195.3 103.3 23.5 98.0 74.5
S6 4 3 − 45.0 118.5 163.5 70.2 175.3 105.1 7.0 90.2 83.1
S7 4 3 − 13.2 414.4 427.6 69.4 147.4 78.0 − 151.3 28.6 179.8
S8 12 3 − 46.3 758.4 804.7 56.0 226.7 170.7 − 415.2 35.2 450.4
S9 4 2 − 10.0 196.8 206.8 95.1 180.2 85.1 10.6 96.5 86.0
S10 3 2.25 − 46.1 18.9 65.0 101.7 209.9 108.2 5.8 56.1 50.2
S11 13 5 − 46.3 758.4 804.7 56.0 226.7 170.7 − 415.2 35.2 450.4
S12 4 4 − 39.3 120.0 159.3 83.7 154.3 70.5 13.0 40.6 27.6
S13 4 7 75.5 84.6 9.1 82.4 152.2 69.8 82.6 95.0 12.4
S14 3 3 − 35.4 − 30.6 4.8 108.9 248.5 139.6 14.5 17.4 2.9
S15 3 4 75.7 101.4 25.7 79.3 125.8 46.5 85.4 98.5 13.1
S16 5 6 − 40.9 108.0 148.9 70.2 173.8 103.6 9.7 77.7 68.1
S17 5 2 66.6 85.3 18.7 75.9 89.9 14.0 80.0 91.9 11.9
S18 3 2.5 74.6 79.3 4.7 68.0 75.4 7.3 96.9 97.7 0.9
S19 4 3.3 − 21.1 152.5 173.6 74.8 123.5 48.8 25.5 96.7 71.2
S20 4 0.33 − 48.8 − 12.5 36.2 166.8 250.5 83.7 3.9 33.6 29.7
S21 3 1.5 − 45.0 − 36.3 8.7 157.2 229.9 72.7 7.0 13.2 6.2
S22 3 3 − 21.8 81.9 103.7 94.9 128.7 33.8 25.1 96.9 71.8
S23 3 0.58 − 44.3 44.6 88.9 95.2 292.3 197.1 7.4 41.6 34.2
S24 4 3 − 47.0 27.7 74.7 112.4 217.1 104.7 5.5 60.7 55.3
S25 3 2 − 48.0 340.6 388.5 116.0 227.8 111.8 − 116.1 29.8 145.9
S26 3 1 − 32.6 − 9.1 23.5 71.5 133.2 61.7 16.9 34.9 18.0
S27 4 1 78.7 99.5 20.7 65.0 119.0 54.0 84.7 99.3 14.6
MIN 3.0 0.3 − 48.8 − 36.3 4.7 56.0 75.4 7.3 − 415.2 13.2 0.9
MAX 13.0 7.0 78.7 758.4 804.7 166.8 292.3 197.1 96.9 99.3 450.4
AVERAGE 4.6 3.0 − 10.4 161.0 175.4 89.1 179.4 92.4 − 24.5 65.0 90.2
STDEV P 2.4 1.6 45.8 199.3 210.3 26.3 53.3 46.2 124.6 30.6 112.7
18 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

Figure 13. Raw data and 3D visual representations of pedagogical generated solution space.

number, and time. However, the two students, S8 and S11, the resulting individual approaches can be divided into
who recorded the largest quantity of iterations, also two broader categories: random and linear. Random explo-
recorded solutions with optimal performance in NPV and ration can be defined here as students who began their
EUI when compared to the results from other students. optimization process by randomly selecting parametric
This suggests that while there is no observable guarantee values with little consideration of given objectives before
that increased iterations yields higher performing results, comparing the results. After a few iterations, students who
there does appear to be an increase in the possibility of took this approach would then proceed to select their high-
identifying higher performing solutions. est performing model based on the available feedback.
Students were also asked to document their exploration The other approach employed was linear based. Students
and decision making process. Of the 27 students observed, using a linear method would begin with analyzing an
Gerber and Lin 19

Figure 14. Illustrates an example of a linear ‘manual’ exploration process as mapped through the pedagogical benchmark case.

initial design alternative of interest before varying para- considered as enabling the design process to explore more
meters with the intent of improving performance regarding design iterations, as there is a measurable decrease in the
the three overall objectives; then, they would analyze their time accrued between manual and automated design alter-
new model, review feedback, and proceed with a new var- native generation. Given Moore’s law we would expect
iation based on the results. Figure 14 provides an illu- this affordance provided by the system too only get better.
strated example of a linear exploration process. In this
case, it is observable that there is a shift in the design
intent after each iteration analysis. For example, after the
5.2. Accuracy
first stage, the intent shifts from maximizing the SPC During the pedagogical experiment set, we observed an
score to adjusting for the improvement in energy perfor- issue regarding the manual transfer of information from
mance. However, it can be noted that these adjustments the parametric model to the provided Excel worksheet calcu-
were heavily reliant on the user’s intuition combined with lator. During the validation process it was discovered that
the previously generated results. As a result, due to the errors were present in approximately 50% of the student
project’s complexity, the fruition of the user’s adjusted recorded data due to calculated values being placed in the
intent was not always realized, as illustrated in Figure 14, wrong fields in the provided Excel worksheet to miscalcu-
through subsequent iterations. We believe this to be an lated values being used as a basis for decision making. One
important observation for further study as it suggests the error in particular that was repeated was the failure to prop-
importance of design cognition memory, e.g. computa- erly associate new floor levels in the parametric model with
tional approaches for managing design complexity. their proper program occupancy type. This resulted in invali-
dated results for all three optimization scores that were then
used to drive design decisions resulting in invalid design
5. Comparative study and summary alternatives. Considering the simplification of this experi-
5.1. Complexity and speed ment, it can be extrapolated that more opportunity for human
based error would be found when applied to real word sce-
As previously observed during the hypothetical cases, the narios possessing significantly more complexity than experi-
run time necessary for analysis and design complexity enced during this bounded experiment. In comparison, by
have a direct correlation. However, in the case of compar- automating this process the opportunity for human based
ing the prototype driven hypothetical scenario 10 with the error is significantly reduced. In this regard, the automated
pedagogy experiment set, this correlation no longer prototype is arguably enabling the design process by provid-
applies. This is due to the design scenarios being identical ing more consistently dependable analysis results for the
and therefore possessing identical levels of complexity. As generated design alternatives.
a result, differences in times recorded can then be attrib-
uted to the human factors of the manual process necessary
to both analyze and implement the desired changes in the 5.3. Solution space comparison
geometry. During the pedagogical experiment, students With regards to the comparison of the solution pool there
averaged 16.4 minutes per iteration compared to the proto- are two areas of interest. First is the quality of the solution
type’s six minutes per iteration. In this regard, the pro- space generated by our methodology using the Beagle
posed methodology and prototype technology can be under the same time constraints as experienced on average
20 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

Table 4. Performance comparison of the solution space for hypothetical scenario 10 between the pedagogical experiment results
and the prototype generated results after a runtime of three hours, seven hours, and six generations.

INITIAL DESIGN SOLUTION SPACE


Pedagogical H.D.S. Beagle
3 HR 7 HR 6th Gen

NPV (Million $) − 40.9 MIN − 47.9 − 40.9 − 40.9 − 40.9


MAX 340.6 432.1 584.1 834.5
Improvement 381.5 473.0 625.5 875.4
EUI (kBtu/sq ft/yr) 173.8 MIN 66.8 58.7 58.7 56
MAX 292.33 233.4 233.4 233.4
Improvement 107.0 115.1 115.1 117.8
SPC 9.7 MIN − 151.27 − 134.4 − 266.3 − 404.4
MAX 99.3 81.2 82.7 87.5
Improvement 89.6 71.5 73.0 77.8
Pareto Solution (%) 3 OBJ 26.9/36.7 22.4/40.0 19.4/37.2
(Pedagogical/H.D.S. Beagle)

during the pedagogical experiment. The second area of design alternatives is highly dependent on a combination
interest is in comparing the quality of the solution space of factors of the individual’s initial design, exploration
generated by the methodology using the Beagle when method, and their exploration strategies. In contrast, the
allowed an extended time period or 6th generation of design use of the Beagle is able to consistently enable the gen-
alternatives. Table 4 provides the performance ranges of eration of higher performance design alternatives and
the solution pools generated by the validated student results also enable the exploration of the full spectrum of poten-
and the results generated by the prototype during the three tial parametric variables without being restricted by indi-
designated time periods. The three hour period was selected vidual preferences.
for comparison against the pedagogical results as this was Overall, the prototype was able to provide a more con-
the average time recorded by the students. sistently higher performing solution pool than the pedago-
According to these results the prototype was able to gical solution pool when given the same exploration time
provide a solution pool with a 26.8% increase in the mea- period. However, the quantity of iterations and generations
sured NPV and a 13.7% reduction in the calculated EUI. available during a specified time period is directly depen-
However, the pedagogical solution pool was able to pro- dent on the capacity of the hardware, software, Internet,
vide a 22.2% more compliant spatial programming set. and simulation engine used. Therefore another point of
This was in part due to instant feedback available regard- interest is to see if there was observable improvement over
ing the SPC while exploring their design options through extended generations as will become available during the
the Revit platform. Overall, when ranked according to same time period with improvements in computational
their performance for all three objectives 36.7% of the affordances. For this purpose, the prototype was asked to
design alternatives generated by the prototype were des- run for the extended time period of seven hours and then
ignated as Pareto optimal solutions while only 26.9% of again to reach six generations. When applied to the
the students generated design alternatives received this hypothetical case study 10 the prototype was able to pro-
designation. It should be noted that the ranked solutions vide 10 iterations per hour. The findings are described in
provided by the students included all 67 results gener- Table 4. As shown in Table 4, the prototype was able to
ated by all 27 students excluding only the previously continue to improve the performance of the solution space
designated invalid results. In addition, it was observed through increased generations. For example, with regards
that there was an uneven distribution of designated to maximizing the NPV, the 6th generation reached $834.5
Pareto solutions among the student generated results. million while the 7 hr solution pool reached only $584.1
When a student was able to produce a highly ranked ini- million and the 3 hr solution pool only reached $432.1
tial design, as shown in Table 3 and Figure 13, then million. Improvements in this fashion are observable
there was typically a more often observed incremental across all three objectives. Therefore, with increases in
improvement over subsequent iterations. This resulted in computational affordances, the measurable improvements
some students being able to produce multiple Pareto on the solution pool provided by the proposed MDO
solutions while others did not produce any. This obser- design methodology within a designated time period will
vation implies that the probability of generating Pareto continue to expand.
Gerber and Lin 21

6. Conclusion and discussion In addition, the proposed methodology using the Beagle
is able to negate the issue regarding human based error
The affordances for incorporation of real world design
observed in the manual calculation of the objective scores.
complexity into our computational design prototype and
It can be extrapolated, then, that with increasingly complex
simulation methodology are discussed through both the
design problems there is a greater potential for human error
enabling and disruptive impact on the early stages of the
and therefore an increased benefit to the automated process
design process. To explore this impact this research com-
available through the proposed simulation methodology
pared two specific simulation processes based on a novel
using the Beagle. This would enable more informed early
MDO framework, one using the automation developed
stage design decision making for even more complicated
through H.D.S. Beagle (hypothetical cases) and the other
design projects.
using the same platform but with human designer driven
One element that defies parametric translation is that of
decision making (pedagogical benchmark cases).
aesthetic preference which differs widely between individ-
Measurements of particular interest included the time nec-
uals. As a result, resources will typically be expended on
essary to generate design options and the quality and quan-
optimizing only potential design solutions that satisfy this
tity of the resulting solution pool through each simulation
first element. As the prototype possesses no aesthetic pre-
process. Through the 12 presented hypothetical scenarios,
ference, equally it possesses no aesthetic prejudice. While
H.D.S Beagle repeatedly demonstrated the ability to pro-
it may spend time analyzing solutions that will ultimately
vide more design alternatives with continually improved
be dismissed by the designer, equally it analyzes solutions
performance for consideration than typically available
potentially overlooked by the designer. The result is a
through conventional methods or when compared against
broader based design solution pool with overall improved
the pedagogical benchmark case. This result was espe-
multi-objective performance levels to enable more
cially noticeable for geometrically complex design prob-
informed design decision making inclusive of a more
lems. Thereby, this research demonstrates that our
expansive simulated aesthetic and formal range. We
framework using H.D.S Beagle successfully improved
believe the work to be a successful demonstration of an
design cycle latency between the design, financial, and
improved simulation technology, one that integrates
energy analysis domain by automating and integrating the
design complexity and improves a designers’ ability to
platform with the design, energy simulation, and financial
manage complexity within the simulation enabled early
models. The analysis for these three domains provides the
stage design decision making process.
means by which to perform an extensive trade-off study
otherwise unattainable at the early stage of design. The
availability of these simulated trade-off studies between Acknowledgements
rapidly generated design alternatives allows the design The authors thank the University of Southern California (USC)
framework to enable more informed design decision mak- Dean of Architecture, Qingyun Ma, and the junior faculty
ing during the early stages of design while incorporating research grant program; Ms Bei ‘‘Penny’’ Pan, our initial lead
arguably more design complexity. software developer; Junwen Chen, Ke Lu, Shitian Shen, and
During the comparison of the pedagogical experiment Yunshan Zhu for their continued software development; Prof
Kensek and her class; research scholars Xinyue Ma and Ryan
with the use of the Beagle in its full automation form, it
Conover; Laura Haymond for her review and participation; and
was observed that when instant feedback was available to
Autodesk, Inc. for their generous support within the IDEA Studio
understand the coupling between parametric manipulation program.
and performance the human driven pedagogical experi-
ment was able to provide design alternatives with superior
SPC calculated scores. This can be observed in the Funding
increased SPC scores where instant SPC levels were pro- The work was in part supported by funding from the USC School
vided through the Revit platform. However, since the calcu- of Architecture Junior Faculty Research Fund and in part by
lations for the other objectives of interest required a Autodesk, Inc.
multistep process, an uneven focus on maximizing the SPC
score instead of all three designated objectives equally was Reference
developed. This result implies that there was a single objec- 1. DOE. Building energy data book. http://buildingsdatabook.
tive optimization approach tendency during the pedagogical eren.doe.gov. (2011, accessed 16 April 2012).
design exploration process where priority was given to indi- 2. Architecture 2030. Energy – buildings consume more energy
vidual objectives separately rather than a true MDO than any other sector. http://architecture2030.org/the_problem/
approach. This situation was not the case with Beagle, problem_energy. (2011, accessed 18 April 2012).
which gave equal value to maximizing all three designated 3. Bogenstätter U. Prediction and optimization of life-cycle costs
objectives among the generated design alternatives. in early design. Build Res Info 2000; 28: 376–386.
22 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

4. Crawley DB, Hand JW, Kummert M et al. Contrasting the 20. Shea K, Aish R and Gourtovaia M. Towards integrated
capabilities of building energy performance simulation pro- performance-driven generative design tools. Automation
grams. Build Environ 2008; 43 (4): 661–673. Construct 2005; 14(2): 253–264.
5. Oxman R. Performance-based design: current practices and 21. Akin Ö. Variants in design cognition. In: Eastman CM,
research issues. Inter Arch Comp 2008; 6(1): 1–17. McCracken WM and Newstetter WC (eds) Design knowing
6. Augenbroe G. Trends in building simulation. Build Environ and learning:cCognition in design education. Oxford:
2002; 37: 891–902. Elsevier Science, 2001, pp. 105–124.
7. Attia S, Hensen JLM, Beltrán L et al. Selection criteria for 22. Shah JJ and Mäntylä M. Parametric and feature-based CAD/
building performance simulation tools: contrasting archi- CAM: concepts, techniques, and applications, 1 ed. New
tects’ and engineers’ needs. Build Perf Sim 2012; 5(3): York, NY: Wiley-Interscience, 1995.
155–169. 23. Gerber DJ. Parametric practices: models for design explora-
8. Radford AD and Gero JS. Tradeoff diagrams for the inte- tion in architecture. Dissertation, Graduate School of Design,
grated design of the physical environment in buildings. Build Harvard University. Cambridge, MA: 2007.
Environ 1980; 15(1): 3–15. 24. Gerber DJ. The parametric affect: computation, innovation
9. Hensen JLM. Towards more effective use of building perfor- and models for design exploration in contemporary architec-
mance simulation in design. In: 7th international conference tural practice. Dissertation, Graduate School of Design,
on design and decision support systems in architecture and Harvard University. Cambridge, MA: 2009.
urban planning, St. Michielsgestel, The Netherlands, 2–5 25. Burry M and Murray Z. Computer aided architectural design
July 2004. using parametric variation and associative geometry. In: 15th
10. Flager F, Welle B, Bansal P et al. Multidisciplinary process eCAADe-conference, Vienna, Austria, 17–20 September,
integration and design optimization of a classroom building. 1997.
Info Tech Construct 2009; 14(38): 595–612. 26. Menges A. Integrative design computation: integrating mate-
11. Welle B, Haymaker J and Rogers Z. ThermalOpt: a metho- rial behaviour and robotic manufacturing processes in com-
dology for automated BIM-based multidisciplinary thermal putational design for performative wood constructions. In:
simulation for use in optimization environments. Build Sim 31st annual conference of the association for computer aided
2011; 4(4): 293–313. design in architecture, Banff, Alberta, 13–16 October, pp.
12. Yi YK and Malkawi AM. Optimizing building form for 72–81.
energy performance based on hierarchical geometry relation. 27. Kilian A. Design exploration through bidirectional modeling
Automation Construct 2009; 18(6): 825–833. of constraints. Dissertation, Department of Architecture,
13. Turrin M, von Buelow P and Stouffs R. Design explorations Massachusetts Institute of Technology. Cambridge, MA: 2006.
of performance driven geometry in architectural design 28. Aish R and Woodbury R. Multi-level interaction in para-
using parametric modeling and genetic algorithms. Adv Engr metric design. In: Butz A, Fisher B, Krüger A et al. (eds)
Informat 2011; 25(4): 656–675. Smart Graphics, Vol. 3638, Berlin / Heidelberg: Springer,
14. Pantelic J, Raphael B and Tham KW. A preference driven 2005, pp. 924–924.
multi-criteria optimization tool for HVAC design and opera- 29. Grobman YJ, Yezioro A and Capeluto IG. Building form
tion. Energy Build 2012; 55: 118–126. generation based on multiple performance envelopes. In:
15. Fesanghary M, Asadi S and Geem ZW. Design of low- 25th PLEA international conference on passive and low
emission and energy-efficient residential buildings using a energy architecture, Dublin, Ireland, 22–24 October.
multi-objective optimization algorithm. Build Environ 2012; 30. Rolvink A, Van de Straat R and Coenders J. Parametric
49: 245–250. structural design and beyond. Inter Arch Comp 2010; 8(3):
16. Hamdy M, Hasan A and Siren K. Applying a multi-objective 319–336.
optimization approach for design of low-emission cost-effec- 31. Kolarevic B and Malkawi AM. Performative architecture:
tive dwellings. Build Environ 2011; 46(1): 109–123. beyond instrumentality. New York: Spon Press, 2005.
17. Gerber DJ and Lin S-HE. Designing-in performance through 32. Hensen J, Djunaedy E, Radošević M et al. Building perfor-
parameterisation, automation, and evolutionary algorithms: mance simulation for better design: some issues and solu-
‘H.D.S. BEAGLE 1.0’. In: Beyond codes and pixels, 17th tions. In: 21st PLEA international conference on passive and
international conference on computer-aided architectural low energy architecture, Eindhoven, The Netherlands, 19–
design research in Asia, Chennai, India, 25–28 April 2012. 22 September 2004, pp. 1185–1190.
18. Gerber DJ, Lin S-HE, Pan BP et al. Design optioneering: 33. Flager F and Haymaker J. A comparison of multidisciplinary
multi-disciplinary design optimization through parameteriza- design, analysis and optimization processes in the building
tion, domain integration and automation of a genetic algo- construction and aerospace industries. In: CIB W78 24th
rithm. In: 2012 symposium on simulation foraArchitecture conference on bringing ITC knowledge to work, Maribor,
and urban design, Orlando, FL, USA, 26–30 March 2012, Slovenia, 26–29 June 2007, pp. 625–630.
pp. 23–30. 34. Aish R and Marsh A. An integrated approach to algorithmic
19. Gerber DJ and Lin S-HE. Synthesizing design performance: design and environmental analysis In: 2011 symposium on
an evolutionary approach to multidisciplinary design search. simulation for architecture and urban design, Boston, MA,
In: 32nd annual conference of the association for computer USA, pp. 61–67.
aided design in architecture, San Francisco, CA, USA, 18– 35. Malkawi AM. Developments in environmental performance
21 October 2012. simulation. Automation Construct 2004; 13(4): 437–445.
Gerber and Lin 23

36. Holzer D, Tengono Y and Downing S. Developing a frame- 53. Eastman CM. Explorations of the cognitive processes in
work for linking design intelligence from multiple profes- design, Department of Computer Science Report, Carnegie
sions in the AEC industry. In: Computer-aided architectural Mellon University, Pittsburgh, PA, 1968.
design futures (CAADFutures) 2007 (eds Dong A, Moere 54. Akin Ö. An exploration of the design process. In: Cross N
AV and Gero JS), 2007, pp. 303–316. The Netherlands: (ed) Developments in design methodology. New York: John
Springer. Wiley, 1984, pp. 189–208.
37. Coello Coello CA, Lamont GB and Van Veldhuisen DA. 55. Cross N, Christiaans H and Dorst K. Analysing design activ-
Evolutionary algorithms for solving multi-objective prob- ity. New York: Wiley, Chichester, 1996.
lems. 2nd ed. New York: Springer, 2007. 56. Gero JS and McNeill T. An approach to the analysis of
38. Poloni C and Pediroda V. GA coupled with computationally design protocols. Design Studies 1998; 19(1): 21–61.
expensive simulations: tools to improve efficiency. In: 57. Kalay YE. Performance-based design. Automation Construct
Quagliarella D, Périaux J, Poloni C et al. (eds) Genetic algo- 1999; 8(4): 395–409.
rithms and evolution strategy in engineering and computer 58. Stoyell JL, Kane G, W. Norman P et al. Analyzing design
science: recent advances and industrial applications. West activities which affect the life-cycle environmental perfor-
Sussex, England: John Wiley & Sons, 1997, pp. 225–243. mance of large made-to-order products. Design Studies
39. Wright JA, Loosemore HA and Farmani R. Optimization of 2001; 22(1): 67–86.
building thermal design and control by multi-criterion 59. Simpson TW, Rosen D, Allen JK et al. Metrics for assessing
genetic algorithm. Energy Build 2002; 34(9): 959–972. design freedom and information certainty in the early stages
40. Magnier L and Haghighat F. Multiobjective optimization of of design. In: 1996 ASME design engineering technical con-
building design using TRNSYS simulations, genetic algo- ference and computer in engineering conference, Irvine,
rithm, and artificial neural network. Build Environ 2010; California, 18–22 August 1996.
45(3): 739–746. 60. Clevenger CM and Haymaker J. Metrics to assess design gui-
41. Keough I and Benjamin D. Multi-objective optimization in dance. Design Studies 2011; 32(5): 431–456.
architectural design. In: 2010 symposium on simulation for 61. Clevenger CM, Haymaker JR and Ehrich A. Design explo-
architecture and urban design, Orlando, FL, USA, pp. 5–12. ration assessment methodology: testing the guidance of
42. Bäck T. Evolutionary algorithms in theory and practice: evo- design processes. Journal of Engineering Design 2013; 24
lution strategies, evolutionary programming, genetic algo- (3): 165–184.
rithms. New York: Oxford University Press, 1996. 62. Atman CJ, Cardella ME, Turns J et al. Comparing freshman
43. Frazer J. An evolutionary architecture. London: Architectural and senior engineering design processes: an in-depth follow-
Association, 1995. up study. Design Studies 2005; 26(4): 325–357.
44. Miles JC, Sisk GM and Moore CJ. The conceptual design of 63. Autodesk. Best practices for conceptual energy analysis,
commercial buildings using a genetic algorithm. Comp http://wikihelp.autodesk.com/Revit/enu/2013/Help. (2012,
Struct 2001; 79(17): 1583–1592. accessed 30 July 2012).
45. Wang W, Zmeureanu R and Rivard H. Applying multi-
objective genetic algorithms in green building design optimi-
zation. Build Environ 2005; 40(11): 1512–1525. Author biographies
46. Wetter M and Wright J. A comparison of deterministic and David Jason Gerber, MArch, MDesS, DDes, was appointed
probabilistic optimization algorithms for nonsmooth as an assistant professor of architecture at USC in 2009. He
simulation-based optimization. Build Environ 2004; 39(8): has since been awarded a courtesy joint appointment at USC’s
989–999. Viterbi School of Engineering. Since joining the USC faculty
47. Holland JH. Adaptation in natural and artificial systems: an he has developed and taught curriculum in the civil and envi-
introductory analysis with applications to biology, control, ronmental engineering graduate sequence and primarily as a
and artificial intelligence. Ann Arbor, MI, USA: A Bradford design studio and design technology professor in architecture
Book, 1992. for both undergraduate and graduate core and electives. Prior
48. Goldberg DE. Genetic algorithms in search, optimization, to joining the USC faculty, he was a full-time faculty member
and machine learning. Reading, Massachusetts, USA: at the Southern California Institute of Architecture from 2006–
Addison-Wesley Publishing Co, 1989. 2009. Dr Gerber has also taught at the University of
49. Gero JS, D’Cruz N and Radford AD. Energy in context: a California, Los Angeles’s school of architecture and urban
multicriteria model for building design. Build Environ 1983; design, the Architectural Association’s Architecture and
18(3): 99–107. Urbanism March (DRL) graduate program as a technical tutor,
50. Tuhus-Dubrow D and Krarti M. Genetic-algorithm based at the Laboratory for Design Media at the École polytechnique
approach to optimize building envelope design for residen- fédérale de Lausanne (EPFL) in Lausanne Switzerland, at
tial buildings. Build Environ 2010; 45(7): 1574–1581. Stanford University’s Civil and Environmental Engineering
51. Znouda E, Ghrab-Morcos N and Hadj-Alouane A. Department as a guest instructor, and at the Tecnologico de
Optimization of mediterranean building design using genetic Monterrey School of Architecture, Mexico. Professionally, he
algorithms. Energy Build 2007; 39(2): 148–153. has worked in architectural practices in the United States,
52. Goldberg DE. A note on boltzmann tournament selection for Europe, India, and Asia, including Zaha Hadid Architects in
genetic algorithms and population-oriented simulated anneal- London, England; Gehry Technologies in Los Angeles; Moshe
ing. Complex Systems 1990; 4: 445–460. Safdie Architects in Massachusetts, and The Steinberg Group
24 Simulation: Transactions of the Society for Modeling and Simulation International 0(0)

Architects in California. Dr Gerber managed research and Taiwan University focusing on civil engineering, she acquired
development for an AEC technology startup company as vice her master’s in construction engineering and management at the
president of innovation and subsequently began consulting for same school. During her school years, Eve Lin gathered a spec-
architecture, engineering, and construction firms on global trum of experience working at various construction and develop-
digital practice, advanced uses of associative parametric design ment companies such as the Century Development Corporation.
tools, building information modeling, simulation and design There she assisted in the development of the Nankang Software
optimization technologies and was subsequently a vice presi- Park by providing the green building evaluation. She also aided
dent for Gehry Technologies. He has since sat on advisory in the subsequent implementation of the ISO9000 standard.
boards for AEC technology startups. Dr Gerber spent two years During her first master’s, Eve discovered her interest in green
as a research fellow at the Massachusetts Institute of building and sustainable design. Thus, she pursued her develop-
Technology’s Media Lab working on smart cities with Bill ment in this field by concentrating on sustainable related research
Mitchell. His research focus is on design and computation and after her graduation. Fortuitously, she had the opportunity to
the development of innovative systems, tools, methods, and expand her studies by pursuing a second master’s through the
theories at that intersection. Dr Gerber currently advises and USC Building Science Program. By concentrating on emerging
co-advises PhD students from architecture and engineering on technology with regard to building energy-simulation, Eve Lin
topics including domain integration, interactivity and system continued to push the boundaries of her contribution to sustain-
intelligence, digital media and fabrication, and multi-objective able design. In particular she focused on using her knowledge to
design optimization and design optioneering. His work is sup- push the implementation of the environmental management sys-
ported by industry grants and by the National Science tem for Watt Hall to acquire a more energy-efficient profile for
Foundation. David Gerber received his undergraduate architec- the prominent building of the School of Architecture at USC.
tural education at the University of California, Berkeley Merging her broad background and knowledge in civil engineer-
(Bachelor of Arts in Architecture, 1996). He completed his ing, construction management, building science, sustainable
first professional degree at the Design Research Laboratory of design, and energy simulation, she believes building information
the Architectural Association in London (Master of modeling technology is the solution for more efficient, higher-
Architecture, 2000), his postprofessional research degree performing, and sustainable building. While current design tool
(Master of Design Studies, 2003) and his doctoral studies and technology has not been mature enough to achieve the goal
(Doctor of Design, June 2007) at the Harvard University she believes in, she continues her research in design technology
Graduate School of Design. and energy performance feedback in the PhD in Architecture pro-
gram at USC. In the future, she hopes she can continue to devote
Shih-Hsin (Eve) Lin, LEED AP BC + D, hails originally from herself to the ever-expanding and ever-important world of build-
Taipei, Taiwan. Spending her undergraduate years at the National ing information technology for sustainable design.

You might also like