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

net/publication/317279203

Project Discover: An Application of Generative Design for Architectural Space

Planning

Conference Paper · May 2017

DOI: 10.22360/simaud.2017.simaud.007

CITATIONS READS
17 1,534

8 authors, including:

Danil Nagy John Locke

Columbia University Columbia University
10 PUBLICATIONS   46 CITATIONS    2 PUBLICATIONS   19 CITATIONS   

SEE PROFILE SEE PROFILE

James Stoddart Lorenzo Villaggi

Autodesk Research Columbia University
5 PUBLICATIONS   38 CITATIONS    6 PUBLICATIONS   21 CITATIONS   

SEE PROFILE SEE PROFILE

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

Quantification of Spatial Experience View project

Bionic Partition View project

All content following this page was uploaded by Danil Nagy on 05 June 2017.

The user has requested enhancement of the downloaded file.

 Project Discover: An application of generative design for
architectural space planning
Danil Nagy1, Damon Lau1, John Locke1, Jim Stoddart1,
Lorenzo Villaggi1, Ray Wang1, Dale Zhao1, David Benjamin1
1
The Living, an Autodesk Studio
New York, NY USA
[email protected]

ABSTRACT 1.1 Parametric design

This paper describes a flexible workflow for generative de- In the past decade, a new type of design software has
sign applied to architectural space planning. We describe this emerged which is fundamentally changing the way designers
workflow through an application for the design of a new of- use computers to develop and refine their designs. Known as
fice space. First, we describe a computational design model parametric design software, these tools allow the designer to
that can create a variety of office layouts including locating not only define a final geometric solution, but to describe the
all necessary programs and people using a small set of input entire system behind how a design is generated. Within this
parameters. We then describe six unique objectives that eval- larger system description, the designer can expose specific
uate each layout based on architectural performance as well parameters, or values that drive different variations of the de-
as worker-specific preferences. Finally, we show the use of sign.
a multi-objective genetic algorithm (MOGA) to search
Although such a model takes more work initially to describe,
through the high-dimensional space of all possible designs,
it offers the designer many advantages. First, the parametric
and describe several visualization tools that can help a de-
approach makes it easy to create variations and custom ad-
signer to navigate through this design space and choose good
aptations of a design. Instead of manually creating multiple
designs. We conclude by discussing the future of such com-
versions for different applications, the designer can expose
putational workflows in design and architecture. Our hope is
the critical parameters that drive different variations and au-
that they go beyond basic automation to create an expanded
tomatically generate different versions by changing those pa-
role for the human designer and a more dynamic and collab-
rameters. Second, a well-structured parametric model is
orative interaction between computer design software and
more adaptable to change in the future. Since it is defined by
human designers in the future.
a series of operations, the design can be easily adapted to
changing conditions instead of rebuilding the model from
Author Keywords
Parametric modeling, simulation, genetic algorithms, multi- scratch each time.
objective optimization, evolutionary design, generative de- Most importantly, the parametric approach allows the de-
sign, architecture signer to think through design solutions in a deeper and more
ACM Classification Keywords dynamic way than possible with traditional methods. In a tra-
I.6.5 SIMULATION AND MODELING - Model Develop- ditional approach, the designer studies the design problem,
ment internalizes all of its constraints and objectives, and then uses
their skill and experience to craft a single design solution, or
1 INTRODUCTION a handful at most. With the parametric approach, the con-
Computers and computer-aided design (CAD) software have straints and goals of the design problem can be directly em-
had a dramatic impact on architectural practice since the bedded within the parametric model, which can then be used
emergence of computers in academia in the 1950s, and espe- to automatically generate a variety of solutions. Instead of
cially since the introduction of personal computing in the designing a single solution, the designer can now think of
1980s. Although early researchers envisioned a wide-rang- designing a multi-dimensional ‘space’ of design. Each di-
ing future interaction between computers and human design- mension of this design space represents one of the critical
ers [10], the first computer tools to be widely adopted by ar- parameters exposed by the parametric model, and each indi-
chitectural designers were computerized versions of tradi- vidual design variation can be found somewhere within this
tional drafting and rendering tools. While they allowed de- hyper-dimensional space.
signers to produce content much faster than with traditional
methods, they did not fundamentally change the process of 1.2 Beyond parametric
design. While the parametric approach has broadened the possibili-
ties of design and pushed the boundaries of human-computer
interaction in the design process, the exploration of the de- applied similar optimization methods to a variety of architec-
sign space is still limited by the abilities of the human de- tural problems. However, their optimization criteria are sim-
signer. Although some parameters may be set by the con- ilarly constrained to well-known and easily simulated physi-
straints found explicitly in the design problem, for the most cal objectives such as structural and environmental perfor-
part the human designer must investigate different design op- mance. In contrast, we propose a more flexible workflow that
tions by manually varying individual parameters and evalu- can accommodate a diversity of optimization criteria, includ-
ating each option using their own criteria and intuition in a ing those dealing directly with how space is used and expe-
way not much different than with traditional design methods. rienced at the occupant level
The concept of generative design, as described in this paper, The quantification of spatial experience has also been ex-
addresses this limitation by tasking a computer with explor- plored by a variety of authors. Hillier, et al. [4] proposed a
ing the design space semi-autonomously, and then reporting variety of analytical tools for studying spatial configurations
back to the designer which options it considers promising for which they called ‘space syntax’. Peponis, et al. [11] extend
further analysis. Because a computer can process infor- this work by proposing a universal method for understanding
mation much quicker than a human, such a system allows a plan topology through linear representation. Turner, et al.
much deeper exploration of complex design spaces. Tradi- [12] propose a view-based ray tracing technique for under-
tionally, such an approach has been used to optimize a given standing and analyzing spatial configurations. While the pro-
model to achieve maximum possible performance based on posed methods can help the designer derive quantitative data
concrete objectives [8]. With a model of sufficient complex- about their designs, they are only offered as tools to aid a
ity, however, a generative design system can also be used to traditional design process. In contrast, we extend these meth-
reveal interesting parts of the design space and discover ods and show how they can be used as measures of spatial
novel design solutions that would otherwise be hidden to the performance to guide an automated optimization process.
human designer.
3 METHODOLOGY
To take advantage of the possibilities of generative design, Our proposed workflow of generative design for architecture
the basic parametric model must be extended in two ways. is organized into four steps: (1) the design of a geometric
First, the model must include concrete metrics by which each model which can create many design variations, (2) the de-
design option can be evaluated. Since the computer does not sign of a series of performance metrics which can be used to
have any inherent intuition about design, the human designer measure the performance of a single design, (3) the explora-
must explicitly describe to the computer how to determine tion of the model’s design space through a MOGA, and (4)
which designs perform better than others. Second, the model the investigation of the resulting design data through statisti-
needs to be connected to a search algorithm that can control cal analysis. Furthermore, we propose this method as only
the input parameters of the model, get feedback from the one component within a broader design process. Thus, there
metrics, and intelligently tune the parameters to find high are several steps that must be taken both ‘before generative
performing designs while also exploring the full possibilities design’ in order to establish a design concept to drive the ge-
of the design space. One of the most promising of these al- ometric model and collect necessary data for the perfor-
gorithms is the multi-objective genetic algorithm (MOGA), mance metrics. Similarly, there are a variety of steps that
which uses principles of evolution to create sequential gen- must be taken ‘after generative design’ in order to achieve
erations of designs and evolve them to contain higher per- other criteria and develop the selected design solution to the
forming designs over time [9]. level of a final constructible design.
The remainder of this paper describes our development of a
custom workflow for generative design specifically geared 3.1 Before generative design
towards architectural space planning, and our application of As with any architectural design project, the process begins
this workflow to the design of a new office space. by studying the design problem, understanding its goals and
constraints, and formulating a vision and concept for the de-
2 RELATED WORK sign. The vision of the project was to create a dynamic and
The application of multi-objective optimization towards highly functional new office space for Autodesk in Toronto.
solving complex mechanical design problems is well-known Some of the constraints included:
in the field of engineering. Marler and Arora [8] provide a 1. The outline of the three floors of an existing new build-
good overview of various applications. However, being con- ing where the office would be located
strained to the goals of engineering problems, these applica-
tions are limited to using only structural performance as op- 2. The programmatic requirements, including specific
timization criteria. numbers of shared amenities such as meeting rooms
Liggett [7] provides a thorough historical overview of auto- 3. Occupation by up to 300 workers
mated methods for space planning in architecture, including
4. Diversity of different departments, project teams, and
the use of genetic search algorithms. Derix [2], Keough and
workstyles that the office needed to accommodate
Benjamin [6], Chronis et al. [1] and Gerber et al. [3] have
Based on the vision of the project and these constraints (the To get information about these preferences we distributed
goals of the project were established in a subsequent step), surveys to all individuals and teams in the office, asking their
we developed an architectural concept around breaking up preferences in terms of which amenities they want to be close
the floorplan into a series of individual ‘neighborhoods’. In to, which other teams or individuals they often work with,
this concept each neighborhood is a work-area for an indi- and the office conditions they prefer. Based on this infor-
vidual department or project team. The neighborhoods are mation, we were ready to construct the generative design
divided by shared amenity spaces, which are contained model that could generate unique design solutions and eval-
within standalone rooms. These rooms create visual variation uate each one based on specific performance metrics.
within the office space as a whole, while providing a degree
of privacy and uniqueness to each neighborhood. 3.2 Geometric model
The first step was to create a geometric model that could de-
Once this basic concept was established, the design problem fine a set of neighborhoods within the two main floors of the
became the arrangement of neighborhoods within the build- office building, position shared amenity zones between
ing floorplan, the location of shared amenity spaces, and the neighborhoods, and then locate specific programs within the
assignment and placement of teams and individual workers amenity zones and individual workers within the neighbor-
in the neighborhoods. In architecture, this type of problem is hoods. To create each individual design, our geometric
known as space planning, and deals with the optimal ar- model applies the following algorithm (see Figure 1):
rangement of programs and spaces within a fixed plan. Be-
cause there are so many possible variations, this type of prob-
1. Locate a seed point for every neighborhood
lem is traditionally difficult to solve for a human designer,
and typically relies heavily on intuition and rules of thumb, 2. Draw neighborhood boundaries based on edges equi-
along with iterative design and testing of a large variety of distant from the neighborhood seeds (similar to a voro-
solutions before finally choosing the best one. Due to the noi diagram)
complexity of this problem, it was actually the subject of one 3. For each neighborhood, choose one of the edges along
of the first applications of computing to architectural design which to place a shared amenity zone
[5]. For us it was the perfect problem to test the possibilities 4. Place shared programs within amenity zones based on
of the generative design process previously described. a greedy fill algorithm
Besides exploring many design options, another advantage 5. Assign teams to neighborhoods, also based on a greedy
of the generative design approach is that we can evaluate de- fill algorithm
signs at a much higher level of detail than possible with tra- 6. Assign people to specific desks in neighborhood based
ditional approaches, including evaluating some aspects of the on list order.
design which are often ignored or abstracted in typical space
planning projects. In this case we wanted to judge each de- To establish the neighborhood seeds, a linear spine is drawn
sign not only on global architectural goals such as maximiz- over the plan and the seeds are distributed evenly along this
ing the amount of light in the space, but also on local goals spine. Then, each seed’s exact location is refined by two in-
having to do with the individual preferences of each of the dividual parameters – the first defines the distance to move
office’s future occupants. along the spine from the initial point, and the second defines
the distance to move away from the spine in the perpendicu- and complex goals, some of which are difficult if not impos-
lar direction. A third unique parameter chooses the edge sible to quantify such as beauty, fairness, quality of space,
along which to place the amenity zone by specifying its nor- elegance, and novelty. To deal with this potential difficulty,
malized distance along the neighborhood boundary. The we divide the set of all possible architectural performance
placement of individual amenity programs, teams, and indi- metrics into three groups:
viduals is not parameterized, but is instead directly deter-
mined according to the geometry of the neighborhood  Those that can be easily quantified and calculated us-
boundaries. ing existing tools (e.g. daylight analysis)

With 15 neighborhoods controlled by 3 unique parameters  Those that can theoretically be quantified but cannot be
each, the model is completely described by 45 unique param- computed using existing tools, for which new computa-
eters. Currently, there are no theories or rules for how many tion tools must be developed (e.g. employee work style
individual parameters a model should contain to ensure that preference and activity hotspots)
a robust search of the design space is both feasible and com-  Those that cannot be quantified and must be addressed
plex enough to create a wide variety of design options. In through other means outside of generative design (e.g.
general, the current best practice is to make this number as beauty)
small as possible, while ensuring that each critical aspect of
the design is controlled by a unique, continuous variable. While this classification addresses the current limitations of
The uniqueness of each parameter is important so that the the generative design workflow, the conclusion of this paper
algorithm can directly control each aspect of the design in- outlines some ideas for future research that suggests machine
dependently while searching for the best combinations. The learning as a way to quantify and evaluate goals that are chal-
continuity of each parameter is important because the algo- lenging to compute using direct calculation. In our case, our
rithm should be able to fine-tune the parameter settings by analysis of the project goals along with discussions with the
predicting future results based on past experiences. If each managers and individual workers yielded six discrete design
setting of a parameter yields completely different results, it metrics to evaluate each design (see Figure 2):
will be far more difficult for the algorithm to search through 1. Adjacency preference, which measures the travel dis-
the design space. tance from each employee to their preferred neighbors
Finally, in order to take advantage of learning within the au- and amenities
tomated search process, the entire model needs to be com- 2. Work style preference, which measures the suitability
pletely deterministic, relying only on the input parameters of an assigned neighborhood’s daylight and distraction
exposed to the algorithm to generate each design. No noise measurements to the assigned team’s surveyed prefer-
or random parameters should be utilized in the geometric ences
model.
3. Buzz, which measures the amount and distribution of
3.3 Design metrics high-activity zones
To allow the search algorithm to automatically measure the 4. Productivity, which measures concentration levels at
performance of each design generated, we also defined a set individual desks based on sight lines to other desks and
of unique goals, or metrics, which rate the relative perfor- other noise sources
mance of each design along a set of criteria. These metrics
form the set of output values that the search algorithm can 5. Daylight, which measures the total amount of natural
use to evaluate how well each design option performs, and to daylight entering the space throughout the year.
guide its search of the design space toward discovering 6. Views to outside, which measures the ratio of work-
higher performing designs. spaces with an unobstructed view to the exterior glass
One apparent limitation of the generative design process is façade
that all performance criteria for a given design system must One of these – daylight – is well understood and can be cal-
be exposed to the search algorithm as a numeric quantity. culated using existing analysis tools. The other five were ei-
Thus, any performance metric that we want the algorithm to ther novel or highly specific to our design goals. For these
consider must be both quantifiable and computable in a reli- we developed our own custom analysis tools which we built
able and efficient way for all solutions within the design directly into the generative design model.
space.
Each new design project potentially brings with it a unique
In engineering applications where similar optimization set of goals and performance requirements, which will never
workflows have been explored for a number of years, the be fully described in any given design software. Thus, part
metrics are relatively straight forward. For example, the of the responsibility of the designer in the generative design
strength of a structural component is easy to compute using workflow is to be able to use computational tools such as
standard finite element analysis (FEA) software. An archi- parametric modeling and custom scripting to describe their
tectural design problem, however, often has many competing
unique design goals to the computer. Although this some- space. Subsequent generations are then produced by either
times makes the design task more difficult, it also has the directly taking high performing designs from the previous
potential to expand the role of the human designer while generation (a process called elitism), or randomly mixing the
opening up new opportunities for design though enhanced parameters of two high performing designs to create a single
human-computer interaction. new design (a process called cross-breeding). Each new de-
sign’s input parameters may also be slightly modified before
Along with the geometric model, the design metrics consti-
it enters the population (a process called mutation). This pro-
tute the second half of the full generative design model. This
cess is then repeated for multiple generations, either until the
model is a closed system that (1) takes in a discrete set of
target number of generations is reached, or performance fails
input parameters, (2) creates a unique design solution based
to improve for a certain number of generations. In this way,
on those parameters, (3) evaluates the design along a set of
a MOGA uses concepts found in natural evolution to gener-
unique metrics, and (4) outputs those metrics as a set of dis-
ate new designs based on the input parameters (genome) of
crete values. When this system is connected to a search algo-
previous high performing designs, thus gradually promoting
rithm, it can be automatically explored for good design solu-
the best options (survival of the fittest) and ‘evolving’ higher
tions. However, although the algorithm can explore many
performing designs over time.
more designs than possible through traditional manual
means, it can only evaluate them based on the specified met- This type of algorithm has many advantages in the context of
rics output by the model. Thus it is crucial that the chosen generative design. As the name implies, the MOGA can op-
metrics sufficiently capture the priorities of the design prob- timize designs along any number of output metrics. Further-
lem, and accurately describe the relative performance of each more, the user does not need to prioritize or weight the indi-
design according to those metrics. vidual metrics beforehand. This is because the MOGA deter-
mines relative performance based on the idea of dominance
3.4 Design evolution rather than the absolute difference in metric values. A design
Once we have defined the generative design model, we can is considered better performing than another if it dominates
use a search algorithm to automatically explore the space of or performs better in one or more of the metrics. Thus the
possible designs and discover novel and high performing de- algorithm will continue to produce designs that are dominant
sign options. A search algorithm is a subset of a general op- in as many of the metrics as possible, and the user can later
timization algorithm, which is concerned with discovering decide how to prioritize the metrics.
optimal settings of input parameters of a function which
maximizes the value of one or more outputs. Although many Another advantage of the MOGA is that it works stochasti-
search algorithms exist, the one of particular interest to us is cally by sampling designs from the design space, and trying
the multi-objective genetic algorithm (MOGA). to learn optimal configurations of the input parameters
through experimentation. Other optimization algorithms
This algorithm generates designs in groups called genera- such as gradient descent rely on computing gradients for
tions. The first generation is composed of a set of initial de- each objective with respect to each input parameter. This is
signs either randomly or evenly sampled from the design not possible with most parametric design models, which are
defined by a large number of geometric functions, none of
3.5 Data analysis
which can be easily differentiated. Thus, such model can This process generated a data set containing 10,000 designs,
only by optimized through a stochastic experimental process. including the input values for each design and its score along
Finally, genetic algorithms have also been shown to be ex- the six metrics. One approach at this stage would be to filter
ceptionally good at finding the overall best performing de- the dataset by the metric scores and directly select a few
signs within a design space (the global optimum) while high-performing designs for further analysis. However, de-
avoiding locally high-performing areas that may not be the pending on the complexity of the design problem such a se-
best overall. By recombining high-performing designs from lection can be challenging for a number of reasons.
different areas of the design space, and slightly mutating de- First, the various metrics might be directly competing with
signs over time, genetic algorithms can avoid local optimums each other, which means that there is actually no single best
more effectively than simpler, more deterministic algorithms design but a range of equally high performing designs along
such as gradient descent. the trade-off between competing metrics. For example, when
As with any optimization algorithm, the MOGA has hyper- designing an industrial component there is typically a trade-
parameters that need to be set before beginning the search off between the part’s weight and its strength. In this case,
process. These hyper-parameters have a significant impact unless there is a specific weight or strength target, it would
on how the algorithm behaves and thus are an important as- be difficult to select a single ‘best’ design without first un-
pect of generating good results. However, these settings also derstanding how this trade-off works.
depend on the nature of the problem, so their tuning is often Second, as previously mentioned, the hyper-parameters of
a product of heuristics and previous experience. The MOGA the MOGA have a significant effect on how the search
hyper-parameters include: works, and proper tuning of these settings depends on the
 The sampling method or the starting population particularities of each generative design model (including
how many and what type of input parameters and output met-
 The size of the starting and subsequent populations
rics are used). Thus, it is rarely enough to run only a single
 The termination criteria of the process (run for a set search process, and it is helpful if the results of every search
number of generations, or continue until no new better are studied in depth to determine how the hyper-parameters
designs are found for a number of generations?) may be tuned for future runs.
 Cross-over rate, which dictates how many of a genera- Finally, one of the advantages of a learning-based process
tion’s designs are created by combining two designs such as MOGA is that it not only finds high-performing de-
from the previous generations signs but also performs the search in a structured, semi-intel-
 Mutation rate, which dictates the rate at which a de- ligent manner. By investigating the search process itself,
sign’s parameters are slightly modified before entering more can be learned about the nature of the problem as a
the next generation whole. In order to investigate this process and gain a deeper
understanding of the design space, we developed a series of
In our case, we used generations of 100 designs each and ran
data analysis tools to aid the designer in exploring the dataset
the process for 100 generations creating 10,000 designs. The
of designs generated by the MOGA.
starting population of 100 designs was generated by ran-
domly sampling from the design space. Through experimen-
Inheritance analysis
tation we settled on settings of 95% for cross-over, and 0.2%
for mutation. The entire process ran over 5 days on a single In addition to the input and output data for each design, the
MacBook Pro with a 2.60GHz Intel Core i7 processor and 16 MOGA also outputs a history of how these designs were gen-
GB RAM. erated. Figure 3 shows a plot of this data, with each point
representing a design, and each column of points represent- input parameters. Once we have assigned the clusters to each
ing a generation of designs. Two colored lines entering a design we can study how these design typologies relate to
point from the left indicates that the design was formed performance in the output metrics. For example we can see
through cross-breeding of those two designs. A thin black if certain design types perform better in some metrics than
line indicates that the design was carried over directly into others. Such tools can help us understand the design problem
the next generation. in general and reveal potential design strategies, rather than
simply picking the single best design.
In this plot you can see an instance where a newly formed
design is high performing and thus is consistently carried
Metric space analysis
over into future generations (A), as well as a case where a
new design gets carried over one generation but then dies out, Once we have understood the distribution of designs in the
likely due to the fact that it was not as high-performing as input space, we can study how the designs perform along the
others in its generation (B). Studying such plots helps us un- six performance metrics. Since there are usually less output
derstand how the algorithm explored the design space, how metrics than input parameters, the space of outputs is not typ-
dominant design lineages are formed, and helps locate poten- ically as high-dimensional as the input design space. Never-
tial blind spots in the design space missed by the algorithm. theless, if there are more than 3 or 4 metrics it can be difficult
to represent the results on a single plot. Our typical approach
Input space analysis and clustering is to do a pairwise plot of all the output metrics to find com-
binations of metrics that have an interesting relationship or a
To analyze how the sampled designs are distributed within
clear trade-off. We can then study the tradeoffs in greater de-
the design space, we can use principal component analysis
tail by plotting them against each other on a scatter plot (see
(PCA) to transform the 45-dimensional input space into a
Figure 5).
new 45-dimensional space where the dimensions are now or-
dered according to the extent to which they describe the var- Once we have studied the performance of the whole set of
iance in the data. Then we can use the first two PCA compo- designs, we can select a subset for further manual analysis.
nents to create the best-possible two-dimensional projection As a baseline the MOGA will provide us with a set of designs
of the high-dimensional design space and see how the sam- which are statistically dominant called the Pareto designs. To
pled designs are organized within that space. narrow it down further we can look for designs that occur at
different points along the trade-offs, which can help us to see
To further study the distribution of designs in the design
the effect of those trade-offs on the design solution. We can
space we can cluster them based on Euclidean distance in the
also use the cluster information generated earlier to identify
full 45-dimensional design space using the K-means algo-
cases where similar performance was achieved by different
rithm (see Figure 4). Intuitively, this gives a representation
typologies of designs.
of different design typologies or strategies that share similar
 3.6 After generative design knowledge of various design factors such as comfort, beauty,
Once a set of interesting designs is selected, they can be fur- or novelty that are crucial to good design but have tradition-
ther analyzed by the human designer, discussed with the ally been difficult to relate to a computer.
stakeholders, and developed into a final design. It is im-
portant to note that since the MOGA follows a stochastic pro- As these types of workflows continue to develop in the fu-
cess based on sampling a limited number of designs from the ture, it is our hope that they not only allow designers to de-
design space, the overall optimal design will not necessarily velop high performing design options, but also help them un-
be found through the search process. Furthermore, as dis- derstand their design problems better through a more collab-
cussed previously, not all aspects that are important to an ar- orative human-machine design interaction. This will allow
chitectural design can necessarily be represented as a metric us to move far beyond the basic automation of tasks evident
in the generative design model. Some aspects, such as in early CAD tools, and leverage the full potential of true
beauty, cannot be quantified, and thus need to be considered computer-aided design.
once the generative design process is complete.
REFERENCES
Finally, most generative design models including the space- 1. Chronis A., Tsigkari M., Giouvanos E., Aish F., Zaki A.
planning model presented in this paper are fairly abstract and A., “Performance driven design and simulation inter-
oversimplified, providing only rough geometry, boundary, faces: A multi-objective parametric optimization pro-
and location information. After a basic space-planning strat- cess”, SimAUD (2012)
egy is selected, there is still much refinement and design 2. Derix, Christian, “In-Between Architecture Computa-
work to be done, including selecting architectural materials tion”. In: International Journal of Architectural Compu-
and designing connection details, to get it to the level of a ting, Issue 04, Volume 07, (2009), 565-585.
final constructible design.
3. Gerber, D., Lin, S., Pan, B., Solmaz, A.S. “Design
Therefore, the process does not end with choosing one of the optioneering: multi-disciplinary design optimization
designs found by the algorithm. Instead, a deep analysis of through parameterization, domain integration and
various high performing designs and their trade-offs should automation of a genetic algorithm,” SimAUD (2012)
suggest potential design strategies that the designer can fur-
4. Hillier, Bill, et al. "Space syntax." Environment and
ther explore to achieve a final best design.
Planning B: Planning and Design 3.2 (1976): 147-185.
4 CONCLUSION 5. Johnson T., Dietz A. G. H., Weinzapfel G., Drauss R.,
This paper described our development of a generative design Morris D., “Space Arrangement,” AD Architectural De-
workflow for architecture, and our application of it for the sign, volume 9 (1969)
design of a new office space for Autodesk in Toronto. 6. Keough I., Benjamin D, “Multi-objective optimization
Although the results of this investigation have been very en- in architectural design,” Proceedings of the 2010 Spring
couraging, the process also has some limitations. Currently, Sim Multiconference, Orlando, Florida (2010), 1-8
the placement of programs and individual people in the plan 7. Liggett, Robin S. "Automated facilities layout: past, pre-
depends on the neighborhood geometry, and thus cannot be sent and future." Automation in construction 9, no. 2
directly controlled by the MOGA. To get a better and more (2000): 197-215.
targeted search we would need to develop methods to di-
8. Marler R.T. and Arora J.S., “Survey of multi-objective
rectly parameterize this placement and expose those param-
optimization methods for engineering,” Structural Mul-
eters to the algorithm.
tidisciplinary Optimation 26 (2004), 369–395
Another limitation is that the calculation of each design is 9. Murata, T. and Ishibuchi, H., “MOGA: multi-objective
still relatively slow – about one minute for each design – genetic algorithms,” in Evolutionary Computation, IEEE
which limits the amount of exploration we can do. Automat- International Conference on (1995), Vol. 1, p. 289
ically analyzing 10,000 designs already dramatically im-
proves the capacity of a human designer, but is relatively 10. Negroponte N., “Through a Humanism Through Ma-
small considering it is sampled from a 45-dimensional design chines,” AD Architectural Design, volume 9 (1969)
space. Distributing the execution of designs within a single 11. Peponis J., Wineman J., Bafna S., Rashid M., Kim S. H.,
generation over several computers in a network would allow “On the generation of linear representations of spatial
many more designs to be evaluated. configuration,” Environment and Planning B: Planning
and Design, 25 (1998), 559-576
Finally, the workflow can be improved by integrating other
types of modelling, particularly machine learning, for quan- 12. Turner A., Doxa M., O’Sullivan D., Penn A., “From iso-
tifying aspects of the designs that are difficult or impossible vists to visibility graphs: a methodology for the analysis
to compute through direct calculation. This is particularly in- of architectural space”, Environment and Planning B:
teresting because it might allow the computer to develop Planning and Design, 28 (2001), 103-121

View publication stats