You are on page 1of 14

Victor Okhoya April 2015

CARNEGIE MELLON UNIVERSITY


DOCTOR OF PROFESSIONAL PRACTICE
RESEARCH PRINCIPLES II - DRAFT PROPOSAL

SPACE PLANNING AND COMPUTATIONAL CREATIVITY


Investigating the use of artificial intelligence to recognize and generate
architectural space plans?

EXECUTIVE SUMMARY
The main aim of this study is to perform a series of tests to verify whether computers can use
machine learning techniques to recognize or generate architectural space plans. The secondary
aim of the study is to relate this question to a broader question from the theory of computational
creativity - can computers be creative? If computers can recognize or generate architectural
space plans is this evidence that they can be creative?

This study will be of interest to both architects and computer scientists. For architects, if
computers can generate viable space plans then they can be a useful tool in design synthesis
allowing architects to evaluate many more design options than they could otherwise. For
computer scientists, this will contribute to a better understanding of the capabilities and
limitations of computation. In particular it will contribute to research in the area of computational
creativity.

Can computers think? This question was asked by Alan Turing and he created a test to answer
it - Turing’s Test. In similar fashion the proposed research will contribute to answering the
question - can computers be creative? We will discuss the relationship between these
questions. Does the fact that a computer can be creative imply therefore that the computer can
think in the sense that Turing intended? And, in any case, how are we to determine whether or
not a computer can be creative?

The creative activity that this study will focus on is architectural space planning. The study is
motivated by the ability of machine learning techniques to perform tasks like visual pattern
recognition. The question then arises - can this capability be extended to the recognition of
architectural space plans? The study is also motivated by the ability of genetic algorithms to
generate music. The question in this latter case is - can this capability be extended to generate
architectural space plans?

It is acknowledged that there have been several other research studies relating to the
automation of space planning in architecture. These studies, and whether or not they have
successfully answered the question ‘can computers be creative’, will be discussed? In particular
we will ask the question of whether automated space planning qualifies as computational
creativity. What makes this study unique is both the approach to the space planning problem as
well as relating space planning to broader questions from the theory of computational creativity.

1
Victor Okhoya April 2015

The method of executing this study will involve two stages. First will be a literature review to
discuss questions related to computational creativity. What is computational creativity and why
is it important? How do we evaluate computational creativity? Does architectural space planning
qualify as a computationally creative activity? And is creative thinking the same as ‘thinking’ in
the way Turing intended?

The second stage will be a series of software simulations on appropriate test data to establish
whether machine learning algorithms can recognize space plans as well as whether they can
generate space plans themselves. Machine learning algorithms will be trained and tested on
encoded space planning data to see if they score highly. What constitutes a ‘high score’ in this
regard will need to be defined. Following this a generative test to establish whether a genetic
algorithm can generate viable floor plans will also be run.

We will conclude with an analysis of the results from the software simulation in light of the
discussion on computational creativity. Does a positive result from our computational space
planning tests provide evidence that computers can be creative? Does it provide evidence that
computers can ‘think’ in the way Turing intended?

SCOPE & JUSTIFICATION


This study will seek to accomplish two things. First, it will seek to answer the question - can a
computer be creative? The specific creative activity with which it will be concerned is
architectural space planning. Second, the study will seek to perform software simulations to
verify whether machine learning algorithms can learn to recognize space plans, and whether
genetic algorithms can generate viable space plans.

Motivations for this study come from, among others, two areas of active research in artificial
intelligence. First, is the use of machine learning for identifying artworks. Blessing & Wen1, for
example, describe the use of Support Vector Machines for classifying art works from different
prominent artists. The question this study asks is - can this be extended so that a machine
learning algorithm can recognize architectural space plans on a design grid? In other words can
a machine learning algorithm learn to distinguish between an architectural space plan and
random configurations of the design grid?

Second is the use of artificial intelligence to generate music. Phon-Amnuasik et al.2, for
example, used artificial neural networks to extract music structure from existing human music
and create new music by an evolutionary algorithm method. The question this study asks is -
can this be extended to extract structure from existing architectural space plans and then use
this structure to guide the generation of viable new space plans using genetic algorithms?

Both these motivating ideas are related to ideas from computational creativity. “Computational
creativity is a subfield of artificial intelligence in which researchers aim to model creative thought
by building programs which can produce ideas and artefacts which are novel, surprising and

1
Blessing, A., Wen, K., 2010.
2
Phon-Amnuaisuk, S., Law, E., 2007.

2
Victor Okhoya April 2015

valuable, either autonomously or in conjunction with humans.”3 This study seeks to relate these
questions from computational creativity to the broader context of the theory of computation. This
is useful because the theory of computation is usually studied from a mathematical or scientific
standpoint. Yet computers clearly have a role to play in the arts generally, and in architectural
design in particular. This study therefore recognizes the need to incorporate computational
creativity into the broader study of computational theory.

The theory of computation is the study of computable functions. It seeks to explore the
questions of what are the different models of computation and what are the capabilities and
limitations of these models.4 An interesting question arising from artificial intelligence related to
the theory of computation is - can computers think?

Turing asked this question in a seminal paper written in 1950.5 In this paper he not only asked
whether computers can think, he asked how we, as human beings, are able to tell whether or
not a computer can think. He devised the Turing Test to answer this question. If a human
interrogator cannot tell whether she is receiving answers from a fellow human or a machine,
then the machine is said to be capable of thinking.

This definition has not been without its critics. Moor6, for example, provides a discourse where
he highlights and attempts to respond to categorical criticisms of the Turing Test. With specific
relation to computational creativity Pease and Colton7 also provide criticism of the Turing Test
as a tool for determining whether or not machines can be creative. In particular they criticize the
Turing Test and variations of it as lacking interaction which they see as key in the interpretation
of the aesthetic value of creative work.

The study will therefore be divided into two parts. The first part of the study will focus on the
theoretical questions. We will introduce computational creativity as an emergent new field of
artificial intelligence and explain why it is of importance. We will then discuss the question of
how to evaluate whether a computer has been creative. Using this creativity criteria we will
examine space planning to see if it qualifies as a creative task. We will also examine if different
approaches to computational space planning are equally creative or whether some are creative
while some are not. Finally we will relate these questions to the Turing Test. If a computer
generates a space plan that we cannot distinguish from what a human agent can produce has
the computer passed the Turing Test? And does this contribute evidence to the claim that
computers can be creative?

In the second part of the study we focus on the experimental aspects. We will test artificial
intelligence algorithms to see if they can recognize and/or generate space plans. Our
knowledge representation approach will consist of a colored design grid on which the different
spaces of our space plan will be represented by different colors (see Table 1.). We will encode
3
Pease, A., Colton, S., 2011.
4
Sipser, M., 2006
5
Turing, A.,1950
6
Moor, J., 1976
7
Pease and Colton, 2011

3
Victor Okhoya April 2015

this design grid into a linear string and classify several such strings to create a training set. We
will then run a battery of machine learning algorithms on the training set to see if any of the
algorithms register remarkably high scores. Finally we will develop and run a genetic algorithm,
inspired by automated music generation algorithms, on encoded design samples to see if the
algorithm can self-generate new viable architectural space plans.

Table 1. Design Grid of colored squares

We note that although there have been many studies on architectural space planning8 and
although there have also been a number of studies on architectural space planning using
artificial intelligence and evolutionary programming methods9, the uniqueness of the proposed
study lies in: the knowledge representation approach that uses an encoded design grid for
machine learning; the use of contemporary machine learning algorithms to try and recognize
encoded architectural space plans; and the use of genetic algorithms to try and generate viable
architectural space plans based on a learned objective function (see the Research Plan
section).

The attempt to relate the findings from an architectural space planning exercise to broader
questions from computational creativity are also unique to the study.

8
Hsu & Krawczyk, 2003.
9
Dutta & Sarthak, 2011.

4
Victor Okhoya April 2015

BACKGROUND RESEARCH
Background research for this study will be divided into two parts – computational creativity and
architectural space planning.

Computational Creativity
Computational creativity will deal with the theoretical questions regarding whether or not
computers can be creative, whether Turing’s Test can be used to determine if computers can be
creative, and how computational creativity relates to the broader question – can computers
think?

Sipser (2006), describes several models of computation together with their limitations. In
particular, he describes the Turing Machine model as the most general form of computation and
describes its limitations. Sipser provides a good discussion of the capabilities and limitations of
computation. The question of whether computers can be said to think is deeply related to the
computation models of the theory of computation.

Turing (1950), asks “can computers think” and tries to identify what thinking in the context of
computation means. He introduces the Turing Test as an experiment to show that a computer
can be considered to be thinking if a human cannot distinguish its response to a question from
that of another human being. The discussion from Turing’s paper will be expanded to ask
whether a computer can be creative. The creative task in question will be architectural space
planning.

Moor (1976), provides an analysis of Turing’s Test in the light of criticisms that it is defective.
Moor argues that Turing’s Test is important if it is interpreted inductively – i.e. if a computer
passes the Turing Test then this does not prove that computers can think but rather adds
evidence to the idea that computers can think. Moor, however, acknowledges important
limitations to the Turing Test. For example the test, in and of itself, does little to guide research
and investigators have often proposed alternatives to it for evaluating computer performance in
artificial intelligence studies. Moor’s essay is important because it sheds light on the criteria that
would qualify a computer as thinking in broader terms than merely passing the Turing Test.

Pease and Colton (2011), focus the question of Turing’s Test on creativity and discuss whether
the test is appropriate for evaluating creativity. They discuss versions of the test that have been
used to measure progress in computational creativity. They argue that the test is inappropriate
for measuring creative effort because it lacks interactivity which they consider as important for
contextualizing the aesthetic value of art work. Their paper is important because it presents the
idea that even if we can experimentally prove that computers can generate architectural space
plans, we will need a more sophisticated method of determining whether this means that
computers can be creative, or can think, than the Turing Test.

Blessing and Wen (2010), describe the use of machine learning for the problem of identifying a
work of art as having been painted by a particular painter. This suggests that machines can be
trained on distinct and discriminative features of art and design. This study is particularly

5
Victor Okhoya April 2015

important because it suggests machine learning can gain insights into “hidden patterns.” By
hidden patterns is meant patterns that are not superficially evident but which apply consistently
to a body of data. Thus although space plans have very different visual configurations they
possess hidden patterns that allow us to recognize viable space plans from random
configurations. Blessing and Wen provide the anticipation that machine learning algorithms will
discover the hidden patterns and recognize viable space plans.

Phon-Amnuaisuk and Law (2007), describe a method of searching for melodic fragments using
self-organizing maps to encode the training melody and genetic algorithms to develop
computational musical compositions that can be recognized by the machine. The proposed
study will attempt to repurpose the methodology described in their paper to search for
architectural space plans that can be recognized and reported by the machine.

Architectural Space Planning


This will review existing research on space planning. The first question will be – does space
planning qualify as a creative task in the same fashion that composing a poem or painting a
portrait are considered creative? We will also ask the question of whether any existing method
of space planning automation can be said to pass the Turing Test. Finally we will discuss the
question of what distinguishes our study from other space planning research.

Hsu and Krawczyk (2003), provides an overview of the different approaches to computational
space planning over the last 30 years. It discusses space planning programs as well as space
planning processes. It also discusses appropriate methods of representation of computational
space planning results. This will be an important spring board to other work in architectural
space planning.

Dutta and Siddhant (2011), survey 16 space planning approaches that have been attempted
and documented using various kinds of evolutionary algorithms. They provide a discussion of
the merits and demerits of evolutionary algorithms for space planning. Their study will be a
useful pointer to other studies where a wealth of ideas on the use of artificial intelligence
methods for the proposed study can be found. Their study also makes clear that no other study
is approaching the problem quite in the same way as the proposed study. The proposed study is
therefore a unique contribution to the ongoing research.

6
Victor Okhoya April 2015

RESEARCH PLAN
The study will be divided into two parts – literature review and software simulation. Below we
describe a series of research questions that will be addressed through the literature review
process. We also describe the steps that will be taken during a software simulation experiment
to establish if machine learning algorithms can recognize space plans, or genetic algorithms can
generate them.

Literature Review
The literature review will seek to answer four questions related to space planning and
computational creativity:

1. What is computational creativity and why is it important?


2. How do we evaluate computational creativity?
3. If a computer can recognize or generate space plans is it computationally creative?
4. Can the Turing Test be applied to space planning?

What is computational creativity and why is it important? (Estimated Time: 1 week)


Since many architects have not heard of the emergent field of computational creativity the study
will begin by introducing and defining this branch of artificial intelligence. We will use Boden
(2010) to define and discuss computational creativity. We will adopt her definition of a creative
idea as one which is novel, surprising and valuable. We will also discuss Boden’s three main
types of creativity – combinational, exploratory and transformational – and discuss space
planning in relation to these types. We will mention Boden’s example of a generative shape
grammar for architectural design as an example of computational creativity applied to
architecture.

We then provide an example of computational creativity in action. We use Varshney et al.


(2013) as an interesting and novel example of computational creativity. They propose to
demonstrate a data-driven approach to computationally creating culinary recipes and menus.
Despite their focus on cuisine the methodology they describe is adaptable to other disciplines
including architecture.

How do we evaluate computational creativity? (Estimated Time: 1 week)


The idea of evaluation is important for this study for two reasons. First, it is important to know
what kind of tasks qualify as creative. For example a poet writing a poem is typically viewed as
being creative while an accountant balancing books is typically not. Second, it is important to
know how we can tell if a product has been produced creatively. Thus, continuing our example,
if a poet writes a particularly bad poem then we do not classify it as creative.

We will need to apply these ideas to our space planning exercise. Does space planning qualify
as a creative task in the same fashion as writing a poem? Or is it more like an accountant
balancing books? If space planning is a creative task how do we know which space plans are
actually creatively achieved and which one are not?

7
Victor Okhoya April 2015

We will use Galanter (2013) to define and discuss creative evaluation. We will begin by defining
aesthetic evaluation and how it applies to computation. We will discuss why computational
aesthetic evaluation is difficult to achieve and what this means for space planning evaluation.
We will then discuss different methodologies for aesthetic evaluation and introduce the idea of a
fitness function as an evaluating criterion. The fitness function will be important when we
perform our experiment on generative space planning using genetic algorithms.

We will then use Brown (2014) to specify the criteria that will be used for aesthetic evaluation.
Brown’s paper discusses appropriate evaluation methods and measures, who the possible
evaluators are and what their knowledge levels might be. Brown then presents a set of
components, constituting an evaluation framework, focusing on actions, knowledge needed and
the context for evaluation. The proposed study will adopt Brown’s method to develop a
framework for evaluating space planning.

If a computer can recognize or generate space plans is it computationally creative?


(Estimated Time: 1 week)
We begin by distinguishing the act of recognition from the act of generation. In our software
simulations we will use machine learning algorithms for recognition while we will use genetic
algorithms for generation. We will reference Galanter (2014) to discuss this distinction in terms
of what he refers to as analytic (recognition) and generative (generation) modes.

We then discuss whether or not space planning qualifies as a creative task. This may seem like
a surprising question but Hanard (2015) identifies activities that do not qualify as creative.
These include: deduction, induction, abduction, learning, problem solving, imitation, heuristics,
and trial and error. It is therefore reasonable to apply evaluating criteria such as we developed
above to assess whether space planning is indeed a creative activity.

Lastly, we note that space planning has several different approaches. Dutta (2011), for
example, identifies five approaches to space planning using evolutionary algorithms: genetic
algorithm, interactive genetic algorithm, multi-objective genetic algorithm, annealed neural
network and fuzzy logic. We therefore need to assess the different approaches to space
planning to see if they all qualify as creative activities using the criteria we developed above.

In conclusion, by the end of this section we will have answered the following questions: Do both
recognition and generation qualify as creative acts of computation? Does space planning in
general qualify as a computationally creative activity? Do all approaches to space planning
qualify as being computationally creative?

Can the Turing Test be applied to space planning? (Estimated Time: 1 week)
We start by describing the Turing Test in detail based on the Imitation Game in Turing (1950).
We then provide a balanced commentary of the Turing Test based on Moor (1975). Finally we
discuss the validity of Turing’s Test in light of John Searle’s Chinese Room.

Turning to the case of creativity we will discuss Boden (2010) who makes the case that the
Turing Test has been passed in several categories of art and creativity by various computer

8
Victor Okhoya April 2015

programs. We will also discuss the criticisms of Pease and Colton (2011) who hold the view that
the Turing Test cannot be applied to creativity.

We argue that space planning is special because it has objective requirements that transcend
human abilities. Thus, it is possible for a human agent to fail the space planning Turing Test. If
we are evaluating a poor space plan it is not obvious whether we are evaluating a failing human
agent or a failing computer. This is not the case in the Imitation Game; all human behavior is, by
definition, human and a human agent cannot fail the Imitation Game no matter how bizarre their
behavior.

Software Simulation
In the table below we describe the steps to be taken for the software experiments to investigate
if a machine learning algorithm can learn to recognize architectural space plans or whether a
genetic algorithm can learn to generate them.

SOFTWARE SIMULATION
TASK DESCRIPTION TIME
Input space plan data  Create a design grid of colored squares in an MS 4 weeks
into design grid Excel file (see Table 1). Save this file as
Design_Master.xls.
 Make 500 copies of this file and number them
sequentially.
 Using sources such as the Sears Catalogue (see Fig
1) create 500 unique floor plans – one in each file’s
design grid. (See Fig 2).

Figure 1. Sears Catalogue house plan sample.

9
Victor Okhoya April 2015

Figure 2. Space plan input on design grid

 Ensure sources are similar and consistent enough to


aid learning.
Encode design grid  Create a macro that encodes each design grid into a
linear string (see Fig 3).

Figure 3. Encode design grid into linear string.


1 week

 Create a macro that extracts all linear strings into a


single training set table. Classify all strings as YES.
 Create a macro to append 500 random string
configurations. Classify all appended strings as NO.
Sort the table (see Fig 5).

Run ML algorithms on  Load the training set data into a machine learning tool 1 week
training set like Weka (see Fig 4).

Figure 4. Weka interface.

10
Victor Okhoya April 2015

 Train and test the following machine learning


algorithms:
o Decision trees
o Naïve Bayes
o Support Vector Machines
 Record correctly classified instances.
 Use percentage splits and cross over validation to
improve accuracy.
 Use non-training data to test predictive ability.
Generate space plans  Adopt the method of Phon-Amnuaisuk et al.10 to use
using genetic genetic algorithms for the generation of space plans.
algorithm  Each space plan is represented as an encoded string.
 A fitness function for evaluating viable space plans
will need to be developed.
4 weeks
 The genetic algorithm crosses over parents and
mutates individuals in its search for viable space plan
individuals.
 The fitness function evaluates proposed individuals
until a viable candidate is found and reported.

... Y Gr Gr Gr G G G G B B Gr Gr Gr Y YES
... Gr Gr Gr Gr Gr Gr G G B B G B B Y NO
... Y Gr Gr B B B Gr G Gr Gr B Y Gr Y NO
... Y G G Y Gr Y Gr Gr Gr Gr G B B Y YES
... Y Gr Gr B B Y Gr G B B Y Gr Gr Gr YES
... Y Gr Gr G G G B B Y Gr Y Gr Gr Gr YES
... G B B Gr G G G G B B Y Gr Gr Gr YES

Figure 5. Encoded training set.

FEASIBILITY & RESULTS


We do not anticipate any challenges with the literature review. While there will be a substantial
analysis of literature required to answer theoretical questions we expect this to go smoothly.

With regard to software simulations, we anticipate that there will be significant effort required to
develop the training set data required for machine learning. Inputting space plans onto the
design grid is an effort that could last up to a month. We estimate that meaningful learning to
take place we will require at least 500 training instances and preferably 1000 training instances.

A second challenge during software simulation is developing a genetic algorithm for the
generation of space plans. In particular, the development of a fitness function will present a
challenging analysis of the abstract properties space plans that can be captured in an algorithm.
It is expected that a substantial amount of trial and error will be involved and the exercise will
last for at least a month.

10
Phon-Amnuaisuk et al., 2007

11
Victor Okhoya April 2015

We expect that machine learning algorithms will perform well on their classification of
unclassified instances of viable space plans. We also expect that a genetic algorithm will be
able to find a viable space plan in its search space within minutes or hours. Finally we expect
that analysis of the relevant literature will enable us to make the case that if computers can
recognize or generate space plans they can be said to design, and if they can design then they
should be considered as thinking objects.

12
Victor Okhoya April 2015

BIBLIOGRAPHY

Blessing, A., Wen, K. (2010). Using machine learning for identification of art paintings. Technical
report, Stanford University.

Boden, M. (2010). The Turing Test and Artistic Creativity. Kybernetes, 39(3), 409-413.

Boden, M. (1998). Creativity and Artificial Intelligence. In Artificial Intelligence 103 (1998) 347 –
356.

Brown, D. (2014). Computational Design Creativity Evaluation. In Design Computing and


Cognition. Springer 2104.

Dutta, K., Sarthak, S. (2011). Architectural space planning using evolutionary computing
approaches: a review. In Artificial Intelligence Review, 2011, Volume 36, Number 4, Page 311.

Eastman, C. (1971). Heuristic algorithms for automated space planning. Institute of Physical
Planning, Carnegie-Mellon University.

Galanter, P. (2012). Computational Aesthetic Evaluation: Past and Future. In McCormack, J.,
d’Inverno, M., (Eds). Computers and Creativity. 255 - 293. Springer.

Hanard, S. (2015). Creativity: Method or Magic? Retrieved on 12 April 2015 from


http://users.ecs.soton.ac.uk/harnad/Papers/Harnad/harnad.creativity.html

Hsu, Y., Krawczyk, R. (2003). New Generation of Computer Aided Design in Space Planning
Methods - a Survey and a Proposal. In Proceedings of the 8th International Conference on
Computer Aided Architectural Design Research in Asia. 101 - 116. ResearchGate.

McCormack, J., d’Inverno, M. (Eds). (2012). Computers and Creativity. Berlin: Springer.

Moor, J. (1976). An Analysis of the Turing Test. In Philosophical Studies, 1976, Volume 30,
Number 4, Page 249

Nassar, K., (2010). New advances in the automated architectural space plan layout problem. In
Computing in Civil and Building Engineering, Proceedings of the International Conference, W.
TIZANI (Editor), 30 June-2 July, Nottingham, UK, Nottingham University Press, Paper 193, p.
385,

Pease, A., & Colton, S. (2011). On impact and evaluation in computational creativity: A
discussion of the Turing Test and an alternative proposal. In Proceedings of the AISB
symposium on AI and Philosophy.

13
Victor Okhoya April 2015

Phon-Amnuaisuk, S., Law, E. (2007). Evolving Music Generation with SOM-Fitness Genetic
Programming. In Giacobini, M. (Ed). Applications of Evolutionary Computing. Berlin: Springer.
557 - 566.

Sipser, M. (2006). Introduction to the Theory of Computation (2nd Ed). Boston: Thomson.

Turing, A. (1950). Computing Machinery and Intelligence. Mind, 49, 433 - 460.

Varshney, L., Pinel, F., Varshney, K., Bhattachariya, D., Schogendorfer, A., Chee, Y. (2013). A
Big Data Approach to Computational Creativity. In Proceedings of the Conference on Human
Factors in Computing Systems.

Wong, S. S., & Chan, K. C. (2009). EvoArch: An evolutionary algorithm for architectural layout
design. In Computer-Aided Design, 41(9), 649-667.

14

You might also like