Sunteți pe pagina 1din 85

Environmental Optimisation Methods in Sustainable Design Process:

In Combination with Evolution-Based Digital Technology

by

Go Kawakita

Submitted to the Department of Architecture


in partial fulfilment of the requirement for the degree of

Master of Science in Energy Efficient and Sustainable Building


at Oxford Brookes University

April 2008

Abstract

This dissertation discusses emergent, sustainable design methods combined with evolutionary
computation and environmental simulation tools. Digital technologies are frequently applied to
engineering fields resulting in improved solutions. On the other hand, very few architects can
benefit from such technical advances in the field of sustainable architecture. The main aim of this
thesis is to research and develop novel design methods corresponding to the production of
environmentally optimised architecture by using digital technologies.

The completed projects examined in the first part of the dissertation are analysed to investigate
possible future developments of digital tools applied to sustainable design. Moreover, the
analyses are expanded to create an experimental design tool that generates and optimises
windows on selected walls under given environmental criterions. The system uses a Genetic
Algorithm as an optimisation method, which is integrated into ECOTECT to provide
environmental simulations and enable the output of more optimal solutions. Finally, some
solutions to the practical problems of implementing this particular design methodology are
discussed.

i
Acknowledgments

I would like to show appreciation to all my advisors. Ben Doherty, who has been my supervisor

during this work, has helped me to learn a lot including technical aspects. Mary Hancock, the

course chair of MSc in Energy Efficient and Sustainable Building, gave me an opportunity to

achieve my special topic.

I also would like to thank everyone who has supported me in this work. Especially, I am very

grateful to my family; my father, my mother and my sister.

ii
iii
Contents

1. Introduction 1

1.1 Background – Architecture and Digital Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Motivation and Main Aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2. Case Study4 4

2.1 Computer Science – Evolutionary algorithms and design . . . . . . . . . . . . . . . . . . . . . . 4

2.1.1 3D Virtual Creatures – Karl Sims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1.2 Genr8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Architecture – Generative design and computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2.1 Evolutionary Architecture - John Frazer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.2.2 The Generative System - Luisa Caldas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.2.3 Light House - Gianni Botsford . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.2.4 Induction Design - Makoto Watanabe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.3 Summary and Analysis - Emergent Technologies and Sustainable Design . . . . . . . 21

3. Responsive Façade Design System (RFDS) 24

3.1 Description of RFDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.2 Genetic Algorithm – Optimisation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2.1 Overview of GAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2.2 Biological Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2.3 Implementations of a Simple Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . 26

3.2.4 Advantages and Disadvantages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

iv
3.3 Testing Simple Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.3.1 Non-Evolutionary Computational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.3.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.3.3 Results and System Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.4 Responsive Façade Design System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.4.1 System Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.4.2 Coding Method – Genotype and Phenotype . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.4.3 Fitness Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.5 Testing the Responsive Façade Design System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.5.1 Optimisation Processes under Simple Conditions . . . . . . . . . . . . . . . . . . . . 40

3.5.2 Responses of User-Defined Priority Levels . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.5.3 Benchmarks of the RFDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.5.4 Experimental Adaptation to Practical Design Process . . . . . . . . . . . . . . . . . 56

4. Future Developments and Possibilities of the RFDS 61

4.1 Other Search Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.2 Additional Environmental Parameters and Multi-Objective Situations . . . . . . . . . . . 61

4.3 Interactive Operations – Aesthetic Design and Environmental Optimisation . . . . . . 62

4.4 Accessibility to the System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4.5 Complex Geometry and Generative Design Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5. Conclusion 65

Bibliography 67

Appendix A 70

v
List of Figures

2-1 3D virtual creatures evolved for competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2-2 3D virtual creatures evolved for swimming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2-3 3D virtual creatures evolved for walking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2-4 3D virtual creatures evolved for jumping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2-5 Diagram of conceptual modification of Evolving Virtual Creatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2-6 Examples of growth model in GENR8 with external forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2-7 Diagram of basic concept and an example of output

derived from building area optimisation programme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2-8 Objects generated by the Building Envelope Design System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2-9 Optimisation solutions for the case of Oporto . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2-10 Annual energy consumption for Oporto simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2-11 Pareto front toward annual energy consumption and greenhouse gas emissions . . . . . . . . . . . . . . . 14

2-12 The Light House and site conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2-13 Voxel sunlight analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2-14 Environmental data in each height and season . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2-15 The Web Frame; physical conditions for form findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2-16 Form findings and human preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2-17 The flow of human evaluation and development of computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2-18 Keiriki 1 - Form generation and structural optimisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3-1 System flow chart of the RFDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3-2 Simple GA flow chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3-3 Simple test model of the RFDS generating circular windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3-4 Program flow chart of the test model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3-5 Differences in the number of nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3-6 Generative process of the test model showing daylight levels at the reference point . . . . . . . . . . . . 32

3-7 Natural daylight level simulation by analysis grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3-8 An image of the RFDS generating pixel-like façade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3-9 RFDS system flow chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3-10 An image of arrangement of reference points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3-11 Composition of a chromosome and coding method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3-12 The fitness function of the RFDS and function-based weighting factors . . . . . . . . . . . . . . . . . . . . . . 39

3-13 Simple explanations of a test implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

vi
3-14 A generated solution and lighting analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3-15 Lighting analysis with four reference points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3-16 Natural lighting level at each reference point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3-17 Fitness scores of the first simulated model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3-18 Fitness scores of the second simulated model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3-19 Lighting analysis at two height levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3-20 Natural lighting analysis with eight reference points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3-21 Natural lighting level at each point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3-22 Arrangement of reference points and target daylight levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3-23 Fitness scores of the simulation with weighting factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3-24 Natural lighting level of eight reference points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3-25 Differences from each target level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3-26 Initial fitness scores and calculation time in different population sizes . . . . . . . . . . . . . . . . . . . . . . . . 54

3-27 Calculation time and chromosome length in different window sizes . . . . . . . . . . . . . . . . . . . . . . . . . . 54

3-28 Generated windows in different window sizes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3-29 Arrangement of rooms and reference points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3-30 Generated optimum window arrangement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3-31 Lighting analysis of the simulation model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3-32 Lighting analysis showing lighting levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3-33 Fitness scores of the simulation model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

3-34 Natural lighting levels of six reference points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

vii
Environmental Optimisation Methods in Sustainable Design Process:
In Combination with Evolution-Based Digital Technology
Chapter 1: Introduction

1.1 Background – Architecture and Digital Technology

Since computers were invented, many aspects of life have been changed. One might be able to
recognise easily such variations by only sweeping one’s room. There might be a computer itself,
and many pieces of electric equipment controlled by computers. Even non-digital items such as
books, clothes, etc, are also related to computers directly or indirectly through manufacture or
distribution etc. Moreover, it might be noticed that by expanding our sights, almost everything in
relation to our lives is based on digital technologies. In other words, our lives might not be
feasible any longer without computers.

As far as the field of architecture is concerned, the situation is similar to other fields as well.
Architecture was regarded as the result of working by hand. However, the situation is no longer
as it was in the past. Drawings and even model making is produced and controlled by computers.
Furthermore, not only simple tasks such as these, but also design itself has come to belong to
the category of the ability of computers. Design processes derived from pieces of architects’
sketches are replaced by computers in the latest digital design. Various architects such as Zaha
Hadid, Ali Rahim, Greg Lynn, Nox, etc incorporate computers into their design as more than
mere drawing tools. Such trends of digital design are more conspicuous amongst younger
architects. Achim Menges, Alisa Andrasek, Aranda/Lasch, and so on are representative of
younger generation in emergent technology and architecture. In their designs, computers are no
longer a part of their design tools or design processes. Such digital technology is their design
itself.

Although computers are well utilised in architectural design at present, the application of digital
technologies is more advanced in the engineering field, especially structural simulation and
design. Structural members are optimised by digital simulation in terms of construction costs, life
cycle costs, energy consumption and even greenhouse gas emissions. Structural optimisation of
trusses is one of the most typical applications. In addition, optimisation of the size and control of
HVAC systems is also one example in another engineering field.

On the other hand, the relationship between digital design and sustainable design in architecture
is far behind compared with other fields. Currently if we use digital technologies for sustainable
design, conventional methods would generally produce better solutions. In other words, the role
of computers is only to simulate objects previously designed by architects. Firstly architecture is
designed, and it is simulated to gather environmental data. Afterward, the architecture is
changed or fixed partially, and is environmentally analysed again. Such tasks are iterated until
the satisfactory completion of the design, or the budget runs out. Generally, sustainable
architecture goes through such design processes at present. Other non-architectural fields

1
frequently use optimisation to produce improved design outcomes; however, this is not the case
in sustainable architectural design. The high performance of modern computing can now be
used advantageously for sustainable design as well as other architectural fields.

Thus, digital technologies are well combined with architecture in various fields except for
environmental design. Moreover, such connections between architecture and digital
technologies might become more intimate in future compared with present circumstances.
Therefore, as computers are effectively utilised in the other architectural fields, environmental
design must incorporate digital technologies in order to remain current.

1.2 Motivation and Main Aim

The energy consumption in the building sector is currently more than 45% of the UK total energy
use. Therefore, energy savings in the building sector are exceedingly important in terms of
overall energy consumption. Currently, there are various ways of energy conservation including
insulation, thermal mass, controlled ventilations, renewable energy sources, and so on.
Research into such elements has resulted in better building performances in terms of energy
efficiency. Meanwhile, these issues are related to other aspects such as embodied energy, use
of natural sources, etc. According to this, it might be also necessary to research the way to
optimise the utilisations of the natural environment itself, including passive heating, passive
cooling, and so on. In this thesis, an emergent method for sustainable architecture, which related
to digital technologies and passive design methods, is investigated and advanced.

The relationship between sustainable design and digital technologies is not well advanced at
present, as referred to in the previous section. However, according to the high performance of
computation in other architectural fields, the combination between digital technologies and
sustainable design might be a new way of saving energy. Additionally, more than 70% to 80% of
the total energy in buildings is consumed by space heating and lighting in both residential and
commercial buildings. Because of that, it could be very efficient and hold great potential for
energy savings to optimise the passive environment in terms of heat gains, heat losses, natural
light, and so on. For instance, unnecessary windows increase heat losses. Moreover, the glare
derived from bad control of the natural light encourages occupants to use artificial lights. It is the
most important to optimise window arrangements in order to balance the penetration of natural
light with heat losses through windows.

Thus, the main aim of this thesis is to investigate and develop the ways and design processes for
effective use of the natural environment by using emergent digital technologies. Furthermore,
the results of research are expected to be extended and applied to new environmental design
systems. Therefore, the production of simple environmental optimisation tools is partial
objectives of this thesis as well.

2
1.3 Thesis Outline

This thesis describes the conceptual new method of sustainable design in terms of
environmental optimisations based on emergent digital technologies. The main aim of chapter 2
is to research digital techniques at the forefront of design, and critically considering their practical
applications. According to the investigation of present situations of digital design, possible future
developments are also discussed in this chapter.

Based on the results of research in chapter 2, some experimental models in terms of


environmental optimisations are implemented in chapter 3. ECOTECT controlled by Lua which is
simple programming language is applied to the models. The results derived from the
implementations are analysed and utilised for future developments of advanced design systems.

In chapter 4, further stages and possibilities based on the models or systems are demonstrated.
Finally, the overall aspects of sustainable design and digital technologies are concluded in
chapter 5.

3
Chapter 2: Case Study

In this chapter, some examples of digital designs are described and analysed in terms of their
present implementation and future possible developments, especially from the point of view of
sustainable architecture and its design processes. Technical aspects of emergent technologies
are mainly discussed in the first section, and, on the other hand, practicalities of such
technological developments are the objective of the next section, focusing on architecture and
computation. Finally, the potential of digital design and sustainability is discussed considering all
aspects analysed in previous sections.

The use of computers is significant and fundamental in the present architectural field. Almost
every aspect such as drawings, simulations, presentations and so on, is computerized.
Especially, activities such as structural analyses might not be feasible without digital
technologies. Meanwhile, computations are not as effectively applied to sustainable architecture,
compared with other fields. For instance, the forefront digital design, complex environmental
simulations, etc. Considering the successful application of digital technologies to other
architectural fields and advantages derived from them, this thesis seeks to prove that such
technologies must be beneficial to sustainability as well. Therefore, it is necessary for further
developments in sustainable design to research and investigate the advantages and
disadvantages of these emergent technologies.

2.1 Computer Science – Evolutionary algorithms and design

2.1.1 3D Virtual Creatures – Karl Sims

Karl Sims is one of the most well-known computer artists using evolutionary computation in their
work. He shows us the attraction of evolutionary algorithms and their possibilities through his
computer animations. One of his works, Evolving Virtual Creatures (1994) is an exposition of
results of his research about 'virtual creatures' produced and evolved under certain situations. In
this work, 3D virtual creatures are expected to become more complex and behave more
interestingly than when they are first created, as an outcome of optimisation, although virtual
creatures are basically dependent on their fitness in the same way as other evolutionary art
works. Creatures are evaluated with a fitness function and higher adaptations are able to survive
under a given simulated environmental conditions such as gravity, ground friction, and so on.

4
However, the work is different from others in the aspect that control systems which are
user-defined are introduced to decide the forms of virtual creatures, as well as the evolution.
Creatures have some nodes labelled as head, body, limb and so on, but not specific functional
components in their biological concepts. Each node and its connection affect the results of
morphologies and creature behaviour. In addition, creature behaviour is controlled by three
effects which are sensors, neurons and effectors. Such control systems of phenotype
morphologies are the reason why in this work, virtual creatures are more complicated and their
behaviours are more interesting than others produced in ordinary evolutionary ways.

Figure 2-1: 3D virtual creatures evolved for competition.


Source: Karl Sims (1994)

Figure 2-3: 3D virtual creatures evolved for walking.


Source: Karl Sims (1994)

Figure 2-2: 3D virtual creatures evolved for swimming. Figure 2-4: 3D virtual creatures evolved for jumping.
Source: Karl Sims (1994) Source: Karl Sims (1994)

5
Fitness functions are one of the most intricate and stimulating elements in genetic algorithms. A
fitness function is a single value derived from a function which explains the aim of experiment,
and it is the key factor which encourages reaching the conclusion of the experiment. For
example, in the works of Karl Sims demonstrated in figure 2-1 to 2-4, walking or swimming
speed, the maximum height of the jump or the distance to the cube after movement toward it in a
fixed time is used for calculation of the fitness scores. These various conditions for evaluation
result in different outputs even if the given environment for experiments is entirely the same. The
relationship between the influence of the fitness function and a given environment is one of the
points for consideration when evaluating the fitness, and vice versa.

On the other hand, the optimum solution under certain environment is not the optimum solution
under different conditions. According to this, it might be easily understood that species develop
in each environment in a different way. Evolving Virtual Creatures clearly demonstrates such
evolutions of creatures and their possibilities in different conditions including gravity, water
resistance, and so on. In fact, the physical features of virtual creatures are different in each
condition or purpose; swimming, walking and jumping. For instance, paddling and tail wagging
creatures are produced by the swimming fitness function and there are shuffling creatures
produced by the walking fitness function. Each creature evolves strategies to exist in its
particular circumstances.

As demonstrated above, 3D virtual creatures are successfully produced by Karl Sims.


Meanwhile, his works could be modified in terms of architecture and its environmental
optimisations as follows. Creatures are interpreted as virtual space, and creature morphology or
behaviour is regarded as physical space or its use. Figure 2-5 shows a conceptual model of
architectural reinterpretation of Karl Sims’ works by the author. In the first stage of an experiment,
genotype virtual space is created, which has architecturally categorized nodes such as windows,
shading devices, and so on. The virtual space forms phenotype space, which is physical 3D
space with some restrictions. Afterwards, generated 3D space behaves and changes under
control systems. That is to say, this virtual space behaviour involves occupant activities
determined by 'neurons', which is one of operators used in Sims’ works, with the values of
sensors such as lighting sensor, temperature sensors, or weather sensor, for example, opening
or closing windows, using shading devices, etc. However, there are naturally physical limitations
in the same way as the Sims’ experiments, for example, whether windows can be opened or not,
or whether shading devices can be fully closed or not. Thus, simulated occupancy activities are
restricted by imposed physical limits. Meanwhile, a fitness function in this modified case might be
energy consumption or CO2 emissions within a given period. Through these processes and
genetic evolutions, optimised virtual spaces are created. This is one of possible ways to
reinterpret the project, “Evolving Virtual Creatures”, into architecture.

6
Virtual Space (Genotype) Main Space
Window
Shading Device
Wall

Space Morphology Space Volume


Window Size
Shading Device Size
Wall Depth

Occupant Activities Sensors Lighting Sensor


Temperature Sensor
Weather Sensor
Neurons Window - Open or not
Shadings - Active or not
Effectors Physical limitation
Materiality

Virtual Space (Phenotype)

Evaluation (Fitness Scores) Energy consumptions Environmental factors


CO2 Emissions Weather U-value
Sunpath G-value
Season Shading factor
Evolution Air tightness

Figure 2-5: Diagram of conceptual modification of Evolving Virtual Creatures.

2.1.2 GENR8

GENR8 is a surface modelling tool using a biological growth mechanism which is based on map
L-Systems. The map L-Systems are one of the L-Systems specified for surface generation. The
main aim of GENR8 is interactive design, so various parameters are available in order to
generate and evolve surfaces. Basically, in the system, parameters can be regarded as
environmental factors, and surfaces are generated with these surrounding influences. Figure 2-6
shows examples of surfaces produced by GENR8 with external influences. In the left case, a
surface is attracted by gravity and disturbed by a sphere in the same way as in the real world.
Meanwhile, in the right case, cylinders magnetize a surface in relation to the distance between a
surface and attractors.

7
Figure 2-6: Examples of growth model in GENR8 with external forces.
Source: O’Reilly and Hemberg (2006)

Users are not able to control the whole system of form finding in other traditional surface
generation tools, because their algorithms to create surfaces are determined beforehand and the
algorithms are designed by system developers. Manipulation is only possible through initial
setup, and the generative process itself is not modifiable. That is to say, although users can
decide initial input as they like, the ways to create objects are fixed. On the other hand, unlike
such conventional design tools, GENR8 encourages users to access every stage of surface
growth and evolutionary computation because in the system, components related to evolutionary
algorithms are also part of the system’s parameters.

Meanwhile, in aesthetic design, one of the most difficult elements about evolutionary algorithms
is a fitness function. Mathematical functions are very useful in the case of optimisation for 'hard'
quantifiable parameters; however, it is difficult to explain human sensibilities in terms of exact
mathematical functions. Therefore, Interactive Evolutionary Computation (IEC) is used for the
fitness function in GENR8. This is the algorithm aimed at predicting user preferences. Fitness
functions are parametric factors in GENR8 as well as the other elements explained above.
According to this, aesthetic sensations of users are reflected in the generated forms, even
though the evolutionary algorithms are carried out inside the system.

Thus, Genr8 is developed considering interactions between users and the system. They are the
most considerable aspect of the system due to the fact that even if one is not an expert in
programming, generative design is made easily available. Generally speaking, designers or
architects prefer intuitive manipulations, so that complicated and intricate processes are likely to
be avoided. Moreover, it might be true that it is commonly not necessary for designers to learn
such complex computer sciences. Emergent design tools are attractive and many people are
likely to use them; however, the deep mechanisms of such technologies tend not to be the users’
focus. Therefore, in terms of development of digital design tools, accessibility of an application to
its users must be one of the most fundamental and notable aspects.

8
As far as the surface generation process itself is concerned, GENR8 is an innovative design tool
as demonstrated above. Nevertheless, no environmental parameters such as lighting factors,
wind dynamics, sunlight and so on, are currently available. Although it is because GENR8 is
created as not an optimisation tool, but an aesthetic surface modelling tool, the tool would be
significantly more useful for responsive design if it was able to be equipped with environmental
parameters.

There are several possibilities for further development of the system. Firstly, reference points to
measure natural light, internal temperature, and so on, could be one of the additional parameters.
The points would have predetermined values as targets, and surfaces would be generated and
modified under those conditions. For instance, if a reference point about natural light is set up in
the scene, a surface would be derived from the calculated annual average luminance by natural
light. If a surface could grow without covering or unveiling a reference point too much, it could be
applied to passive solar design. The second possibility is the wind dynamics parameter. What
this parameter would do is to simulate and calculate air flow through a generated object.
According to this parameter, ideal forms would be created and optimal air flow made available.
Such technologies might be useful for architecture in the tropics. Thus, if it is not necessary to
consider the technical problems of software development, possible improvements are infinite.

Needless to say, there are various problems special to environmental design software, which
cause difficulties of system developments. From the computer scientific points of view,
complexity of environmental simulation might be the most problematic topic. Due to the fact that
environmental simulations take a great deal of time, the systems are not interactive at all. In
other words, changes of objects in the scene do not respond to the result of environmental
calculations simultaneously. In the same way as rendering in 3D visualization software,
calculations are necessary every time the scenes are changed. These facts might be
troublesome obstructions for the early stage of design tasks or responsive design. Users cannot
work comfortably, for example by sketching their ideas. Moreover, developments of new
software solving such problems take enormous cost and time to produce; therefore, it could be
stated that it is too difficult for individuals to cope with these complexities.

2.2 Architecture – Generative design and computation

2.2.1 Evolutionary Architecture - John Frazer

John Frazer broke new ground in the field of digital architecture. His focus differs from other
algorithmic architecture which creates the latest attractive objects, which might be the
mainstream in present digital design. As far as evolution is concerned, such present
mainstreams are probably mere complicated form finding exercises. In short, they are created by
uncountable iterations of works, whether they are intricate or simple tasks. On the other hand,
Frazer’s architecture is a solution corresponding to certain conditions. They are generated and
evolved in a similar way to the natural world.

9
In his projects, environmental conditions, especially solar constraints, are one of the most
important factors. The Groningen Experiment (1995), which is a computational urban design, is
one of his notable research projects. In the project, he demonstrates a responsive design
process which is similar to the ways that some architects use computational methods at present.
Although natural light is the only factor of generative design at this point, it could be said that his
idea of the relationship between architecture and the natural environment is important due to its
innovative parts. Afterwards, the idea of environmentally responsive design developed with the
growth of technology. In 2003, he put forward a more advanced and complex method which is a
computer programme to generate and arrange architecture, optimised under the maximum
building area and control of solar gains. In short, this system is basically applied to urban design
according to the fact that buildings in certain area are comprehensively optimised. Figure 2-7 is
an example of the output of the system.

Figure 2-7: Diagram of basic concept and an example of output derived from building area optimisation program.
Source: Frazer (2003)

One of important points of the project is the way to interactively arrange generated architecture.
The shadow derived from each building is calculated through a year, and those data are used for
optimisation of layout scheme. The surrounding environment influences the place and size of a
building and vice versa. Needless to say, the project is basically about optimising building
arrangement, and therefore, it is not possible to apply the project to the case of a single building
simulation. However, unlike the ordinary ways of passive environmental simulations in which the
environment influences the result in a unidirectional way, the generative processes in his project
are interesting and remarkable due to the fact that influences are bidirectional between the
surroundings and the generated architecture. Such interaction might be exceedingly important
for sustainable design.

However, the computer program he developed can handle only simple objects such as a cuboid;
therefore, practicality and accuracy are not well achieved in the present circumstances. In
addition to this, not only natural light, but also other environmental factors are necessary for
further development, and other elements such as energy consumption, greenhouse gas
emissions and so on can be applied to such objectives as well.

10
Frazer also shows us many other interesting points in generative design, as well as responsive
design demonstrated above. The Building Envelope Design System is a software system to
generate novel forms by evolutionary computation as figure 2-8 shows. Preliminary outlines and
some supportive axes determined by users create complex forms through several computations.
In terms of the flow of the generative design, the system is considerably developed, and there
are many applicable ideas to other generative systems or generative design processes.
Meanwhile, due to the fact that the main focus of the system is the mathematical form generation
and evolution, created shapes are not composed of any environmental elements. Briefly, they
are simply aesthetic, and not generated with influences by any external conditions.

Figure 2-8: Objects generated by the Building Envelope Design System.


Source: Frazer (2003)

As demonstrated above, each aspect such as generative design or responsive design is


gradually, but definitely being developed, although there are still problems to be solved for
practical use. In the environmental design, one of the most important factors is accuracy or
practicality derived from the placing and orientation of buildings, materials, sizes of each building
part, etc. In fact, many complicated equations are used to simulate environmental conditions in
the project demonstrated above to arrange architecture in terms of maximum solar gain, even if
each model is simple in the project. Such complexity might be the reason why technology
development in terms of environmental design is not mainstream at present. In other words,
designers or architects might prefer to create aesthetic objects, rather than solving difficult
environmental situations and making complicated equations.

11
2.2.2 The Generative System - Luisa Caldas

Luisa Caldas is one of the pioneers who applied evolutionary algorithms to environmental design.
The Generative System (GS) which is a GA-based simulation system invented by her, helps
designers or architects to find solutions mainly in terms of energy efficiency. In a simulation using
GS, there are two main factors to measure fitness scores; annual energy consumption and
satisfaction of design constraints. Annual energy consumption is calculated due to both lighting
and heating or cooling energy, and with consideration of surrounding influences such as solar
gains, heat losses, internal radiation, and so on. On the other hand, design constraints are
slightly more complicated. According to Caldas, GS is not only an evaluation tool, but also a
design tool; therefore, the results of simulations need to reflect the opinion of designers. In other
words, design constraints are part of the fitness function to avoid unexpected results and control
the generative process to a certain extent.

Figure 2-9: Optimisation solutions for the case of Oporto.


Source: Caldas (2001)

GS is applied to existing architecture to compare present energy performance and simulated


solutions. The selected building for the experiment is the school of architecture in Oporto
designed by Alvaro Siza. Figure 2-9 demonstrates the results provided by GS. The window size
and the depth of shading devices are optimised. The red lines are user determined design
constraints in terms of Siza’s design concept. Figure 2-10 shows annual energy consumption for
both existing building and solutions. According to the figure, the reduction of energy consumption
is about 10% for the best solution. Moreover, Caldas (2001) claimed that the reduction might

12
become more, which could be confirmed by a more accurate energy calculation in GS. Thus, the
Generative System encourages finding environmentally better design solutions within
architectural design intentions. This user-dominant optimisation for GA is remarkable, and it can
be developed and applied to advanced generative design tools for further stages.

120 Others
Vent fans
100 Space cool
Space heat
80 Lights

60
oporto-shading 87.58
40 oporto-best 89.99
oporto-average 96.22
20
oporto-existent 96.45
oporto-worst 110.55
0
best-shading best average existent worst

Figure 2-10: Annual energy consumption for Oporto simulation.


Source: Caldas (2001)

However, it is also true that GS currently optimises just window size and the depth of shading
devices. Additionally, the fitness is simply a measure of energy consumption. It is practically
necessary to consider other factors including construction costs, embodied energy, lifecycle cost,
and so on. The lowest energy consumption is not always the best solution for other elements,
and vice versa. For instance, the latest products including glazing might be expensive even
though their environmental performance is better than the older solutions. Such a situation is
known as Pareto efficiency, which is the condition that if some in a group are improved, others
become worse. As far as such multi-objective optimisation problems are concerned,
multi-objective genetic algorithms (MOGA) are applied to GS for the second stage. This is a
substantial approach in terms of pragmatic solutions. The black squares in figure 2-11 explain
trade-offs for two objectives, the reduction of greenhouse gas emissions and annual energy
consumptions, so that users are able to choose more ideal and practical solutions than those for
single functions. Furthermore, in the experiments of GS for Pareto optima, other combinations of
objectives, for example construction cost, energy balance, etc, are also successfully simulated.

13
7000

6000

5000

4000

3000

2000

1000
gen 1
gen 200
0
22 22 24 25 26 27 28 29 30
MWh

Figure 2-11: Pareto front toward annual energy consumption and greenhouse gas emissions.
Source: Caldas (2001)

Thus, because of its environmental optimisations, the Generative System is an interesting and
useful tool even if it is under development. On the other hand, it is a fact that some problems still
remain to be solved. Firstly, at this stage, GS is not integrated into any CAD software, and it is
not developed as a plug-in tool. Solutions derived from GS are not visualized until evaluations
and simulations are completed. This means that intuitive manipulations, which designers or
architects might prefer in the design process, are not available. Therefore, GS does not reach
the level of a generative design tool, and is still a simulation support tool. Secondly, form finding
or form generation is not supported in the tool unlike Genr8. Although shape generation is
experimentally developed as the third stage of GS at present, feasibility is extremely limited and
it is possible to modify only the given rectangular box.

Nevertheless, these limitations under 3D visualization and productivity might not be the most
problematic in terms of the advances of other morphology tools. Integration of technologies in
different research studies is necessary and possible for further improvement. The most serious
problem is differences between virtual and real space in the software. Building fabrics, purpose,
and other physical situations are fundamental for environmental building simulations. For
example, it is impossible to calculate heat losses without U-values, and solar gains without
g-values or glazing types. Moreover, it is not possible to compare building performance to
benchmarks without occupancy. On the other hand, virtual spaces in software, of course except
for environmental simulation tools, do not have any physical existences. A box is not a building

14
or a room. A vertical object is not a concrete wall or a wooden column. Such basic factors for
environmental simulations are obstacles to intuitive manipulations, and are the most problematic
and complicated elements for further software development.

As explained above, GS is an interesting system, although it has various problems to be solved,


including design ability, reality in a virtual space, etc. Especially, approaches toward
multi-objective optimisation problems are notable in terms of comprehensive solutions.
Therefore, it could be claimed that the concepts of the Generative System are useful for not only
development of other tools, but also environmental design itself.

2.2.3 Light House - Gianni Botsford

The Light House in Notting Hill, London designed by Gianni Botsford is one of a small number of
real computational design projects. Due to the restricted site situations, the project is
conspicuous by its use of sophisticated environmental analysis and optimisation. As figure 2-12
demonstrates, the site is surrounded on all sides by three to five storey houses and trees;
therefore, the only way to gain sunlight is through the roof, which Botsford calls the roof façade.
One of basic requirements for the project is to provide natural light to the whole area in the
house; not only the top floor, but also the bottom floors.

Figure 2-12: The Light House and site conditions.


Source: Gianni Botsford (2006)

15
Figure 2-13: Voxel sunlight analysis.
Source: Gianni Botsford (2006)

Figure 2-14: Environmental data in each height and season.


Source: Gianni Botsford (2006)

In the solution to such limited conditions, there are two main important key words; “Solar Logic”
and “Building Envelope Optimisation”. At the first stage of the project, extremely detailed
environmental data was collected by using the Solar Logic which Botsford developed in order to
visualize and analyse the behaviour of natural light. Due to the system, environmental conditions
are demonstrated as a three-dimensional grid by voxels which are 1m cubes. Figure 2-13 and
2-14 are examples of voxel data analysis. The way of explaining natural light by the Solar Logic
can greatly encourage users to analyse existing conditions of the site at any point and any time
through a year. For example, it is possible to recognize three-dimensionally the brightest or the

16
darkest part of the site, or the comprehensive changes of natural light. It is just like the way that
engineers collect and dissect microscopic information for further developments. There is no
entirely subjective design process because of preconceptions. Scientific environmental data is
gathered at the first stage of the design processes. This might be one of the most important
factors for environmental design, although various buildings are controlled by designers’
preoccupations from the first stage of projects. Moreover, it might avoid architects to stray into
time-wasting on unnecessary simulations caused by wrong intuitions.

Although the system explained above is exceedingly interesting and substantial for data analysis,
they are currently visualisations of environmental data, not architecture itself yet. In terms of
architecture, the collected data is used for spatiality and materiality with clients’ preferences
including room arrangement, room size and so on. Meanwhile, as explained above, the roof
façade is the only possible way to obtain natural light in the site; therefore, building materials for
the roof are optimised by using the software developed by Botsford and Arup. The software uses
a genetic algorithm and can be applied to multi-objective optimisation problems. This topic is
also discussed in the previous section, and might be important with regard to practical solutions.
In the case of the Light House, Pareto front between daylight factor and direct sun visibility is
evaluated with glazing types such as opaque, filmed or clear. Due to the many objectives and
therefore multiple fitness functions, reconciling these becomes difficult. Is the most important
factor economical, environmental, or functional optimisation? There could be as many fitness
functions as the number of problems to be solved.

Considerably, in the Light House project, computational design and technologies are actualized
unlike many other projects which remain in a conceptual state. On the other hand, such a real
project demonstrates some remarkable aspects for reality and computational optimisation.
Generally speaking, clients truly control each project. Solutions derived from computers might
not be the best solutions or even satisfactory solutions for them. This is probably more
conspicuous in the case of private projects than public ones. For instance, even if a limited
glazing area on the north façade is a better solution in terms of energy efficiency, one might
require a glass box. Logical solutions could sometimes be against pure preferences.

To sum up, it might mean that due to the fact that architecture always exists with human beings,
it is also necessary for computation to handle humanity. This is deeply related to two huge
topics: multi-objective optimisation problems and interactive evolutionary computation (IEC). IEC
is the algorithm which handles human preferences and evaluations. It might be true that
evaluation of scientific matters is not truly complicated in regards to mathematical functions.
Meanwhile, numerical expressions of human preferences are quite difficult. What is “good”?
What does “good” mean? Computers are very bad at such subjective elements. Moreover, even
if numerical expressions of human preferences were possible, it is not equivalent to the ability to
create preferable solutions. One can select and grade favourite colours, shapes, materials, and
so on; however, a combination of them does not become one’s preference in a high probability.

17
2.2.4 Induction Design - Makoto Watanabe

Induction Design is a series of projects which Makoto Watanabe launched in 1990 and has
researched up to the present. Every project in the Induction Design series is developed by
computer programs to find solutions which adapt to given conditions. He claims that computers
and architecture created by them are not the result of high standard graphic technologies, but
solution based technology. In other words, simplification of works, for example drawing plans, or
random generation of infinite patterns as form controls is not the essence of computation of
architecture. The enormous amount of thought is the true substance of digital design. Therefore,
development of design processes which generate solutions under certain statements, or such
processes for their own sake are the most important element in the Induction Design series.

Some of his digital designs are actualized, although countless other projects by other people
focusing on the computation of architecture, which are similar to his projects, result in virtual
schemes in computer screens. The Web Frame is the first real work of these to demonstrate to
us various aspects of computation in relation to the combination of the real and the virtual. In the
project, the most important factor is the fitness function for the form generative process. Basically,
such value evaluations might be categorized into two main groups in the Web Frame project:
physical conditions and aesthetic sensations. Physical conditions might be regarded as similar
elements which are commonly simulated in other projects, and form generations are executed
under such physical constraints as spatiality, materiality, legal matters, and so on.

Figure 2-15: The Web Frame, physical conditions for form findings.
Source: Makoto Watanabe (2000)

On the other hand, the method of handling aesthetic sensations is extremely remarkable.
Designers’ or architects’ preferences are reflected in fitness measurements in the same way as
physical conditions. However, human sensations are always problematic in computer
programming due to the fact that numerical formulations of human sensations are extremely
difficult. What is the best for somebody? What is aesthetics? In the Web Frame project, such
problems are resolved by assigning scores towards solutions derived from computers.

18
Generated forms by computers are evaluated by the designers, and the fitness scores are used
in the next simulation for better solutions. After iteration of these operations, computers might
become able to satisfy human preferences.

Such concept of computation and humanity is advanced in another project, the Program of Flow.
Figure 2-16 and 2-17 explain the design flow and human evaluation. The important development
in this project is the way to grade human preferences. In the Web Frame, designers or architects
give scores to computers after form generation; meanwhile, in the case of the Program of Flow, it
is possible to decide design intentions in advance by giving hand sketches and their grades. In
short, design constraint is slightly possible. The ability to actively predict user preferences allows
the system to 'design' new forms which already fit the conditions that the system has been
'trained' with. Although use preferences are also considered in Genr8, the way it is executed in
this project is more flexible and intuitive in terms of using hand sketches for design grades, which
many designers might prefer rather than parametrical constraints.

Figure 2-16: Form findings and human preferences.


Source: Makoto Watanabe (2005)

19
Figure 2-17: The flow of human evaluation and development of computation.
Source: Makoto Watanabe (2005)

Another important aspect in the Induction Design series is optimisation. Although this concept is
not achieved at this point in the Web Frame, in projects Keiriki 1&2, the program for structurally
optimised shape generation has been developed and advanced since 2005. Briefly, the software
creates objects with initial parameters and simple models, and the generated forms are
structurally optimised afterward as figure 2-18 demonstrates. One of the most remarkable factors
in the Keiriki projects is materiality and economic efficiency. In the program, structural members
are automatically selected by computers in terms of load, and simultaneously optimised for
economic efficiency. Ornaments, which are not related to structure, and excessive materials are
reduced and removed. In short, the total volume of members is approximately minimized, and
this fact is important not only structurally, but also environmentally according to energy
consumption, including embodied energy. Additionally, Keiriki has advantages in the aspect of
intuitive manipulations. Arbitrary modification of generated shapes is also possible, and the
software finds optimised solutions that correspond to such alternatives.

Figure 2-18: Keiriki 1 - Form generation and structural optimisation..


Source: Makoto Watanabe (2005)

20
Thus, the Induction Design series has achieved and realized various important and remarkable
elements in digital design. On the other hand, problems to be solved also remain. Firstly, as far
as creativity is concerned, algorithms used for form generation in the Induction Design series are
generative, but not evolutionary, even if produced shapes are solutions to given conditions.
Various optimised spider-web-like objects are created in every software execution; however,
other shapes are by no means generated by the software. Needless to say, the extent of
arbitrariness in design is another massive academic topic. In addition to such problems related
to creativity, multi-objective situations are another element. This topic is also discussed in the
previous section, and it means that optimisation under multi-functional conditions has not been
solved in the present circumstances, and is a substantial problem currently. Finally,
manageability as a tool is problematic. The software created by Watanabe requires professional
knowledge of structure, and it is completely different from formal knowledge to operate software
including CAD, 3D visualization software, and so on. In other words, in relation to multi-objective
optimisation problems, deeply comprehensive professional knowledge would be required if
generative optimisation software, similar to Keiriki, which covers every condition including
structure, construction, environment, and so on, were invented. Therefore there are, of course,
too many prerequisites and it would not be suitable for all users.

2.3 Summary and Analysis - Emergent Technologies and Sustainable Design

Many aspects of emergent technology and digital design have been discussed in the previous
sections with completed projects. Since computers became fundamental elements, digital
technologies seem to have been gradually developed. Some systems derived from computers
can imitate even biological mechanisms. The Evolving Virtual Creatures (1994), by Karl Sims is
one of the most remarkable examples, demonstrating such digital innovations as explained in
the previous section. It might also be claimed that computers are no longer simply extensions of
our handwork, but even an augmentation to our brains. As far as digital design in the
architectural field is concerned, generative design tools such as Genr8 or software invented by
John Frazer demonstrate design abilities of computers as creative tools.

Although there are many remarkable aspects of digital technologies, the most significant
advantage of computation might be the ability to process a huge amount of information in a short
time. This is an exceedingly simple task, but a significantly important element. The human brain
might not be able to calculate a huge amount of complicated factors, for instance, environmental
simulations as Botsford has done in the Light House project. Application to solutions of
optimisation problems is also one of examples of such an advantage derived from digital
technologies. In fact, the project in Oporto by Luisa Caldas proves that computers can create
better results than human beings under certain conditions. In terms of sustainability, especially
energy consumption, the fact is remarkable for future development.

21
However, it is also true that such accurate iterations of enormous calculations become obstacles.
Generally speaking, computers iterate their tasks until they finish them, once commands have
been put in, even if the tasks are not finite, and moreover, they calculate all tasks in the same
way. On the other hand, human beings can judge the possibilities of finishing calculations, or
choose the way to calculate. Such judgements are developed by their experiences and intuitions.
Computers do not have such abilities, and the only way to achieve them is well-designed
algorithms created by designers. Thus, digital technologies are not feasible for everyone if they
are not good at computing. Complexities and accessibilities to emergent technologies might be
one of the most serious disadvantages of digital design. In other words, such forefront
technologies are still for designers with specific technology skills.

As explained above, there are some advantages and disadvantages in digital design itself.
Meanwhile, in terms of the further development of computation in sustainable architecture and its
design processes, several problems also need to be solved. Firstly, multi-objective situations are
one of the most problematic aspects. The perfectly optimised solution which satisfies every
objective cannot exist in environmental design. As demonstrated in the previous section, if
architecture is optimised in terms of a certain condition, other purposes might not be achieved
very well. This is deeply related to sustainable design according to the fact that practicalities
such as spatiality, materiality, cost, performance, and so on are the elements which determine
the environmental performance of architecture. Every aspect is actually linked to each other, and
it is impossible to pursue only one element. In the projects of Luisa Caldas or Gianni Botsford,
they search for solutions of multi-objective optimisation problems by using the latest computer
technologies. However, they are not completely solved, and still one of the most complicated
topics.

User preference is the second matter in terms of not only environmental optimisation, but also
design aspects of architecture. It might be easy to achieve targets only numerically. For example,
the purpose to gain maximum sunlight is simply solved with the largest window or glass box. The
best solution for the lowest heat losses is to avoid arranging any windows on the wall. However,
these solutions are, needless to say, not acceptable for designers or architects. Environmental
solutions produced by computers are only useful when they are satisfied with these solutions.
Moreover, projects are more complicated due to the existence of clients. Therefore, optimised
solutions are necessary to be generated in accordance with user preferences. In terms of such
conditions, interesting technology called the Interactive Evolutionary Computation (IEC) is
applied to Genr8 and the projects of Makoto Watanabe. The system handles user preferences,
and reflects them into outputs derived from computers. On the other hand, Luisa Caldas controls
design constraints successfully by using a different technology in her project in Oporto. These
technologies are emergent, and still under development; therefore, further research is
fundamentally necessary.

22
Finally, the way to explain a real space in a digital space might be regarded as the most
problematic factor in the developments of emergent technologies and sustainable design. In a
virtual space of 3D visualisation software, drawn objects do not have any meanings, and they
are the brief output of binary data on a screen. On the other hand, users regard such objects as
buildings, rooms, parts of architecture, etc. This gap of recognition between computers and
human beings is one of the factors that obstruct combinations of digital and sustainable design.
As also explained above in this section, accuracies and practicalities including spatiality,
materiality and so on are exceedingly significant factors for environmental simulations. It is
impossible to simulate environmental performance in any case without these elements. In fact, a
great deal of data must be determined for calculations in environmental simulation software like
ECOTECT, TAS, etc to produce meaningful results.

Such tasks in environmental tools might be acceptable in terms of their main aim. Meanwhile,
they might not be suitable for design tools, because designers or architects generally prefer
intuitive design processes. Therefore, specifications of material, zones, etc, seem to be
inconvenient as early stages of design processes. This is one of the most important reasons why
environmental simulations have not been connected with generative design. The Generative
System created by Luisa Caldas is one of the tools trying to deal with such complexity; however,
successful results have not been achieved yet.

23
Chapter 3: Responsive Façade Design System (RFDS)

In this chapter, the Responsive Façade Design System (RFDS), which optimises window
arrangement considering natural daylight level, is demonstrated. The system is a combination of
evolutionary computation and environmental simulation. One of the technologies of evolutionary
computation, the genetic algorithm, is applied to the RFDS in order to investigate optimum
solutions under certain conditions. In the first part of this chapter, genetic algorithms are briefly
explained including basic concepts, terms, implementations, and so on. Secondly, a simple
environmental design model which is not based on evolutionary computation is tested in terms of
further developments and potential of the system. The final part of this chapter is composed of
the implementation of the RFDS and some results of experimental simulations demonstrating its
advantages and disadvantages.

3.1 Description of RFDS

The RFDS is the optimised window generation tool


considering natural daylight level at user defined Genetic Algorithm
reference points. In the present circumstances, the Optimisation method
system works only in ECOTECT which is widely
used environmental simulation package. By using
the RFDS, it is no longer necessary to iterate
traditional top-down search methods. In Ecotect
conventional ways, designers are required to Environmental Simulation
repeat design and environmental simulations until
they obtain their target, which might not be an
efficient method. On the other hand, the RFDS No
encourages users to find some optimum solutions
Fitness
once initial parameters have been inputted.
Various parameters are available for the
optimisation processes, so that it is possible to
generate windows on any walls and in any sizes. Yes
The detailed implementation and design ability of
the RFDS is demonstrated in the following sections
of this chapter.
Final Result

Figure 3-1 is the flow chart of the RFDS describing


two main components; a genetic algorithm (GA) Figure 3-1: System flow chart of the RFDS
and ECOTECT. In the present system, Lua, which
is a simple programming language, is utilised to

24
control the whole operation in terms of facilitating the system development and accessibility to
ECOTECT. A genetic algorithm is one of search methods, and is currently applied to various
kinds of optimisation problems. At the first stage of the design process, some parameters
defined by users are passed into the GA generating an initial statement for the optimisation
process. Afterward, the GA goes through some computational steps, and the results achieved
by these steps are environmentally evaluated in ECOTECT. If the evaluations satisfy user
targets, executions are finalised and optimised solutions are outputted for user evaluation.
Meanwhile, the same operations are iterated if the evaluations are not adequate for the targets.
The RFDS attempts to generate environmental optimal solutions.

3.2 The Genetic Algorithm – Optimisation Method

3.2.1 Overview of GAs

Genetic algorithms (GAs) are search methods which draw heavily on the metaphor of the
mechanisms of biological evolution. In brief, it is the simulation of virtual creatures in a computer
deriving the creature in a group best adapted to the environment. Due to their exceedingly high
performance in optimisation problems, the algorithms are utilised in various fields including
economics, social systems, scheduling systems, and so on. The list of applications might not be
by any means exhausted.

Originally, GAs were invented and developed by John Holland, his colleagues, and his students
in the 1960s. In contrast to the present widespread applications to problem solutions, the initial
research of GAs was to simply study adaptation methods of natural selection and natural
genetics. On the other hand, it is generally stated that creatures are evolved adjusting to the
surrounding environment by iterating crossover, mutation, and selection. In other words, the
higher the fitness of individuals toward their environment is, the higher the probability that they
will survive and produce offspring is. Such an adaptation mechanism is a problem solution itself,
and the reason or basic concept for the various applications in different fields, although the
original purpose of GAs by John Holland is different from the present utilisations.

3.2.2 Biological Terminology

The following is some of the biological terminology which is used in GA implementations. In


terms of GA implementations, it is necessary to understand technical terms at this point.

Individual: An autonomous piece characterised by a chromosome. In this case, one possible


solution to the design problem
Population: A group of individuals
Population Size: The number of individuals in a population
Gene: A functional block of DNA

25
Allele: A possible value of a gene
Chromosome: Strings of DNA. In this case, a list of parameters
Locus: The place of a gene in a chromosome
Genotype: Genetic explanation of a chromosome (e.g. binary strings)
Phenotype: The physical manifestation of the genotype

3.2.3 Implementations of a Simple Genetic Algorithm

As far as the simplest GAs are concerned, there Start


are just three types of operators; Selection,
Crossover, and Mutation. Figure 3-2 explains the
simplest GA system flow. After some parameters
are initialised, the three GA operators are iterated
until the results derived from the algorithms satisfy Initialisation
terminal criteria defined by users. Thus, GAs are
executed by two very simple elements which are a
loop of GA operators and terminal criterion. Each Evaluation
step of the algorithms is explained in detail as
follows:
Selection

Crossover
Initialisation
In this step, some parameters including
population size, number of generations, Mutation
chromosome length, and so on are inputted.
Afterwards, the initial input randomly
generates genotype individuals of the first
Evaluation
generation. Especially, population size is
significant in terms of the operations that,
No
generally speaking, the longer the
Terminal
chromosome length is, the bigger the Criterion
population size is. Additionally, the bigger
population size requires longer calculation
time until convergence. However, small
Yes
population sizes may result in premature
convergence.
End

Figure 3-2: Simple GA flow chart

26
Evaluation
Fitness scores are calculated for further selection of fitter chromosomes. One of the most
important aspects in this step is the Fitness Function which calculates the fitness
measurement of each individual. This operation is deeply related to the efficiency of the
whole GA flow; therefore, it needs to be determined carefully.

Selection
The fitter chromosomes in the population are basically selected for reproduction. As in
biological evolution, the fitter chromosomes are more likely to be selected and reproduced
in each generation. Meanwhile, lower fitness chromosomes are also possibly selected, but
with a lower probability. This probabilistic selection depends on the selection method.
There are several types of selection such as elite selection, roulette selection, tournament
selection, etc. Each selection type has advantages and disadvantages. For instance, in
elite selection, the fitter chromosomes are certainly selected in order; however, premature
convergence is highly possible. In the RFDS, a combination of elite selection and roulette
selection is used.

Crossover
Crossover roughly mimics the genetic operation of biological recombination between two
chromosomes. The fitter chromosomes are chosen by the selection operator; however, it
is not effective enough to evolve the population. The crossover operator encourages more
variation by exchanging genes between two chromosomes.

Mutation
The mutation operator randomly flips or changes genes in a chromosome between alleles,
generally with a very low probability. Chromosomes generated by the crossover operator
are basically copies of the parent chromosomes; therefore, premature convergence
possibly occurs. Chromosomes that have been mutated help to avoid premature
convergence. Generally speaking, the mutation rate should be 1/L, where L is the length of
chromosome. Moreover, if the mutation rate is too big, the algorithm becomes similar to a
random search.

27
Terminal Criterion
In this step, the conditions required to terminate the GA is evaluated. If the process is
regarded as being completed, the fittest individual in the generation is outputted as one of
the possible optimum solutions. The conditions of convergence are as follows:

- if the fittest score in the population satisfies the certain target – star gene
- if the average fitness score in the population satisfies the certain target – population
improvement
- if the increase or decrease of fitness scores in the population becomes below a certain
value - convergence
- if the number of generations becomes over the defined value – finite iteration

3.2.4 Advantages and Disadvantages

GAs can be characterised by some features such as coded parameters, global heuristic search,
fitness-based selection, and so on. Genetic algorithms are a very general search method, and
are not specific to any particular application. GAs are very effective at avoiding becoming
trapped in local optima. On the other hand, there are also some disadvantages as well. GAs are
fitness-based, and it is necessary to calculate fitness scores for each individual every
generation; therefore, processing loads are very high. In order to produce good solutions, bigger
population sizes are generally necessary, and the bigger population size causes higher
processing loads. Meanwhile, the most serious disadvantage is the complexity and variety of
initial parameters. Satisfactory solutions require accurate input of initial setup parameters such
as population size, mutation rate and so on; however, they are changeable depending on the
objectives. Therefore, it is very complicated and unpredictable for users to determine accurate
setup parameters.

28
3.3 Testing Simple Model

3.3.1 Non-Evolutionary Computational Model

In this section, a simple test model of RFDS is demonstrated. The RFDS is the
evolutionary-based design system finding optimum solutions under user-defined environmental
conditions. However, in this simple test, evolutionary computation and optimum simulation
methods are not yet operated. The main purpose of this simple test is to observe and analyse a
generative and responsive process of the system. Therefore, some aspects of the test model
such as parameters, outputs, and so on are slightly different from the developed RFDS in further
stages. Meanwhile, the main aim of the thesis, that is to investigate and develop the ways and
design processes for effective natural environment by using emergent digital technologies, is
consistent, and is reflected in this model as well. Figure 3-3 demonstrates an example output
generated by the test model of the RFDS generating circular windows.

Figure 3-3: Simple test model of the RFDS generating circular windows.

29
3.3.2 Implementation

The test model is simply implemented as


figure 3-4 explains. At the first operation, Start
some parameters are inputted to initialise
the system, for example, the shape of
window, the number of windows, and so
on. After initialisation, new windows are
generated, and the size of each window is Initialisation
also increased until the environmental
target value is satisfied. The evaluation of
convergence is manipulated after every Generate Window

operation in terms of accuracy toward


target values. The following is detailed
information of each parameter for system Terminal
Criterion
initialisation.

No

Target Daylight Level [lux]


In the system, the terminal criterion is
natural daylight level [lux]. Users are No

allowed to set up the target value of Terminal


Criterion
daylight level at a reference point in
the model. It is possible for users to Yes
place the reference point anywhere in
the model. In the case of this test
model, only one reference point is End
allowed to be arranged in the scene.

Radius of circular windows


New windows are generated in terms Figure 3-4: Program flow chart of the test model
of radius determined by users. The
size of generated windows is
increased constantly in relation to the
initial radius; therefore, the value of
the initial radius is one of the most
important elements affecting the final
output.

30
Maximum number of windows
It is possible to determine the maximum number of windows which will be generated in the
model. Once the number of windows in the scene is beyond this value, new windows are not
generated, and only window size is increased instead in order to satisfy the target daylight
value.

Number of nodes in a window


This is a slightly technical term demonstrating the number of points composing a window.
Basically, the test model generates approximately circular windows as figure 3-3 shows;
however, the shape of windows actually depends on the number of nodes. Figure 3-5
explains the types of window shapes in the differences of the number of nodes. The more
nodes there are, the smoother the window shape is. Meanwhile, the large number of nodes
causes heavy processing loads, and it is necessary for users to consider the whole system
flow at the same time.

Figure 3-5: Differences in the number of nodes

Thus, the simple test model is composed of only four parameters. In addition to this, the
evaluation of convergence is exceedingly simple, compared to other environmental simulation
tools. Therefore, it is true that accurate simulations and complex design are not allowed in the
system. Figure 3-6 is an example demonstrating the generative process of the test model with
the value of target daylight value.

31
Figure 3-6: Generative process of the test model showing daylight levels at the reference point

32
3.3.3 Results and System Developments

Figure 3-7: Natural daylight level simulation by analysis grid.

Figure 3-7 is the daylight analysis of the example generated by the simple test model. In this
case, the reference point is placed in the centre of the room at a height of 600 mm from the floor.
The target daylight level is 700 lux. As the figure shows, the system successfully generated
several circular windows until the terminal criterion is satisfied. On the other hand, in this model,
only one reference point is available, and it is not accurate enough to control the internal
environment. In fact, the example result demonstrates that although the daylight level around the
reference point is nearly 700 lux, the other areas are not under specific control. Especially, the
lux levels of half of the part in the room are above the target level. They might have an
exceedingly high daylight level, and might not be desired in terms of glare.

Meanwhile, due to the fact that the test model is created simply to experiment and investigate
the generative and responsive processes of the system related to natural daylight level, other
environmental and practical parameters are not operated. In terms of the further development of

33
the system, other parameters are necessary to simulate and control more complex situations.
For example, temperature could be applied to one of the other terminal criteria. Moreover,
energy consumption and greenhouse gas emissions might be necessary to be implemented in
future models. It is also important to note that such target criteria require more complicated
logical system because practical values including the physical properties of materials are
fundamental for environmental simulations.

34
3.4 Responsive Façade Design System

Figure 3-8: An image of the RFDS generating pixel-like façade.

3.4.1 System Implementation

As mentioned in the previous section of this chapter, the RFDS attempt to generate optimum
window arrangements under user-defined environmental conditions by implementing a genetic
algorithm within ECOTECT. Figure 3-9 explains the detailed implementation of the system. At
the first point, users are required to input two important parameters for optimisation processes,
which are target daylight levels at each reference point, and the priority of each target.
Afterwards, initial parameters for GA operations need to be given by users. The GA of the RFDS
is a relatively conventional GA, and the algorithm is as explained in the previous section. Once
the system obtains all of required initial parameters from users, possible solutions under the
given conditions are generated, and ultimately produced as a 3D model in the scene at the final
stage. In the RFDS, the terminal criterion is the number of generations. That is to say, once the
programme has performed a given number of generations, programme executions are cancelled
to output solutions. In addition to the final 3D model, the RFDS can output the fittest solution in a
generation. Numerical results are also outputted as an Excel file format. The detailed
explanations of each element related to the system are as follows:

35
Start

Generate Initial Group

Crossover

Mutation

Evaluation

Selection

No

Terminal
Criterion

GA main loop

Yes

Output the Result Wmf

Excel

End

Figure 3-9: RFDS system flow chart.

36
a) Reference Point
Users are allowed to place reference
points for measurement anywhere in a
scene. There are no limitations to the
number of reference points; however, it is
important for users to consider that the
more the reference points are placed, the
more difficult it is for the system to find
optimum solutions.

b) Target Daylight Level [lux]


The fitness scores are calculated in
Figure 3-10: An image of arrangement of reference points
terms of target daylight levels determined
by users. In the RFDS, target levels are
not listed, and are completely user-defined values. Daylight level is demonstrated in lux due to
the fact that, at present, various benchmarks are published related to lighting levels. Therefore,
users are able to easily decide target levels at each reference point.

c) Priority Level of Target


Priority level is one of the elements influencing fitness measurements as well as target daylight
levels. There are three levels from low priority to high priority, and users are required to give
such levels to each reference point. According to this, the system generates possible solutions
giving more important reference points higher priority.

d) Window Size and Interval


In this implementation of the RFDS, windows are arranged in lines as figure 3-8 demonstrates,
so that output is a pixel-like façade. Users are able to change window size and interval between
windows.

e) GA Operators and Initial Inputs


It is necessary for users to input some parameters to initialise the GS within the RFDS. Similar to
other GAs, population size, number of generations, number of elites, and mutation rate are
required as initial inputs. As demonstrated in the previous section, there are three types of GA
operators; selection, crossover, and mutation. In the GA of the RFDS, elite selection and roulette
selection are combined. Uniform crossover is utilised in the RFDS, and occurs to all parent
chromosomes. Uniform crossover is one of the crossovers to generate child chromosomes by
randomly mixing genes of parent chromosomes. Meanwhile, mutation is operated according to a
given mutation rate.

37
f) Output formats of optimum solutions and simulated data
The RFDS outputs solutions as 3D models in a scene; therefore, it is possible to export them to
CAD and 3D visualization software. Additionally, the fittest solutions in each generation are also
outputted as a wmf file which is a vector format image. Numerical data including fitness scores in
each generation, simulated daylight levels at each reference point and so on are exported to
Excel. These data encourage users to analyse not only environmental performance, but also
system performance for future developments.

3.4.2 Coding Method – Genotype and Phenotype

One of the most important elements in GAs is the coding method. In the RFDS, coding and
decoding is operated as figure 3-11 demonstrates. Before running GA, the system generates
virtual grids on walls using the window size and interval defined by users. Alleles used in the
system are binary; 0 and 1. Therefore, a chromosome might be 01001010101101…, and the set
of binary codes of a chromosome is also a set of binary codes of each façade. If each grid
explained above is regarded as genes, it is possible to arrange windows on walls. In the present
system, “1” means a window and “0” means not a window. Due to such definition of phenotype
and genotype, windows are arranged on each façade as the figure shows.

It is also necessary to consider the length of chromosomes before the system proceeds. A
chromosome is simply a set of binary codes, and the coding method is also very simple;
therefore, the smaller window size and interval require larger number of binary codes for a
chromosome resulting in an increase of calculation loads.

Figure 3-11: Composition of a chromosome and coding method.

38
3.4.3 Fitness Function

An experimental method is applied to the fitness function of the RFDS in terms of design abilities
of the system in relation to user preferences. A fitness score is simply a total value of the fitness
measurements of each reference point. The system is an optimisation tool considering target
daylight levels; therefore, in the case of the RFDS, the chromosome of minimum differences
from target levels is the fittest in a population. In other words, in this system, fitness functions
attempt to approach 0. Although the total of each difference from target levels at each reference
point is clearly a possible fitness function, the concept of weighting factor is utilised in the
present model.

As figure 3-12 demonstrates, weighting factors are calculated based on priorities of each target
daylight level. The higher the priority is, the bigger the increase rate of weighting factor is. The
fitness function of the system is the sum of the weighted differences, which are the calculated as
the differences multiplied by weighting factor. It means that the higher the priority is, the bigger
the fitness score is likely to be. According to this, the system is expected to find solutions giving
priorities through GA operations.

Figure 3-12: The fitness function of the RFDS and function-based weighting factors.

39
3.5 Testing the Responsive Façade Design System

3.5.1 Optimisation Processes under Simple Conditions

Some experimental simulations were completed by using the RFDS to evaluate the present
performance of the system, and investigate the advantages and disadvantages for future
developments. The implementations of the first test model including the room and window size
are as figure 3-13 shows. In this model, four reference points without priority are placed, and
each target daylight level is as follows:

Reference point – 1: 200 [lux] – priority Low


Reference point – 2: 600 [lux] – priority Low
Reference point – 3: 800 [lux] – priority Low
Reference point – 4: 400 [lux] – priority Low

The GA parameters are the population size = 20, number of generations = 400, number of
individuals for elite selection = 2, and the mutation rate = 0.03. Additionally, the chromosome
length under the initialised window size is 800 according to the system. All priorities have been
set to low to ensure that the weighting does not bias results.

200
3,000

200

h = 700
100
Elevation Window size

6,000

1,500 3,000 1,500


6,000

Plan Perspective

Figure 3-13: Simple explanations of a test implementation.

40
Figure 3-14 and 3-15 demonstrate the generated optimum solution and the natural lighting
analysis. If reference points are categorised into two groups which are the darker (points 1 and
4) and the brighter (points 2 and 3) in terms of the target lighting levels, the figures show that the
RFDS successfully approximated an optimum solution under given conditions. In other words,
the left area of the analysis grid in figure 3-15 is darker than its right area. As lighting levels at
points 2 and 3 are about 640 lux and 830 lux, and it can be stated that target lighting levels are
sufficiently achieved. As figure 3-16 also shows a relatively successful achievement, lighting
levels at points 2 and 3 approach the each target as the simulation proceeds.

Lighting levels at reference points 1 and 4 are not satisfactory values. The target level of point 1
is 200 lux and that of point 4 is 400 lux, while actual lighting levels are, due to the result, about
380 lux and 570 lux. Although figure 3-17 shows that overall fitness scores are decreased in
each generation, lighting levels at points 1 and 4 are stable after the 150th generation. There are
two possible reasons for such system idleness. Firstly, windows are arranged on all four walls of
the simulated model; therefore, it is difficult to achieve low lighting levels. Secondly, reference
points are placed closely obstructing differences of lighting levels.

41
Tested result under 4 reference points

Lux

Point - 1
Target : 200
379 .06 4

Point - 2
Target : 600
Actual : 642 .72 7

Point - 4
Target : 400
Actual : 566 .08 2

Point - 3
Target : 800
Actual : 828 .17 6
Lighting level analysis grid

Figure 3-14: A generated solution and lighting analysis.

42
Lux

Reference Point - 1 Reference Point - 2


379.064 [lux] 642.727 [lux]

Reference Point - 4 Reference Point - 3


566.082 828.176

Figure 3-15: Lighting analysis with four reference points.

43
Generation

Figure 3-16: Natural lighting level at each reference point.

Generation

Figure 3-17: Fitness scores of the first simulated model.

44
The first experiment shows some interesting aspects of the system. More complicated conditions
are applied to the same model used for the first simulation to investigate further behaviours of
the system. In the first simulation, four reference points are placed in the model, while the
number of reference points is increased to eight in the second simulation, and each point is
arranged as figure 3-19 shows. The following is each target level at reference points:

Reference point – 1: 200 [lux] – priority Low Reference point – 5: 400 [lux] – priority Low
Reference point – 2: 600 [lux] – priority Low Reference point – 6: 800 [lux] – priority Low
Reference point – 3: 800 [lux] – priority Low Reference point – 7: 600 [lux] – priority Low
Reference point – 4: 400 [lux] – priority Low Reference point – 8: 200 [lux] – priority Low

In the second simulation, reference points can be categorised into two groups in the same way
as the first simulation model. Points 1, 4, 5, and 8 are the darker group and points 2, 3, 6, and 7
are the brighter group. According to this, it is expected that the half zone of the room, where the
brighter group is placed, would gain more natural daylight than that of the darker group. As far as
values of each point are concerned, the darker group should achieve lower lighting levels than
the brighter group. However, figure 3-20 shows spatially no clear differences of lighting levels
contrary to the first simulation. Although fitness scores are constantly improved as figure 3-18
demonstrates, daylight levels at each reference point do not successfully approach each target
level according to figure 3-21.

It is because reference points are placed too densely so that it is physically difficult to achieve
such a condition. Additionally, the system is not good at searching for better solutions at a small
range because GAs are blind search methods.

Figure 3-18: Fitness scores of the second simulated model.

45
Point - 5
Target : 400
Lighting level analysis grid Actual : 614.662
height = 1,600
Point - 6
Target : 800
Actual : 741.087

Point - 8
Target : 200
Actual : 472.428

Point - 7
Target : 600
Actual : 741.243
Point - 1
Target : 200

Lighting level analysis grid


height = 700 Point - 2
Target : 600
Actual :658.807

Point - 4
Target : 400
Actual : 528.859

Point - 3
Target : 800
Actual : 744.232

Figure 3-19: Lighting analysis at two height levels.

46
Lighting level analysis grid
height = 1,600

Lighting level analysis grid


height = 700

Figure 3-20: Natural lighting analysis with eight reference points.

47
Figure 3-21: Natural lighting level at each point.

48
3.5.2 Responses of User-Defined Priority Levels

Figure 3-22: Arrangement of reference points and target daylight levels.

Weighting factor for fitness function is experimentally equipped in the RFDS to reflect user
preferences in the optimisation processes. The mechanism of the factor is as explained in the
previous section, and users are able to define priorities at each target lighting level. In this
section, results derived from an experiment in terms of weighting factor are discussed to
investigate the possibilities of the factor and its future developments. In this simulation, eight
reference points with priority levels are arranged, and figure 3-22 demonstrates priorities and
target lighting levels of eight placed points in the model. There are two groups with three priority
levels; 700 lux and 300 lux.

The system finds better solutions in each generation as figure 3-23 demonstrates in a similar
way to other simulations. Lighting levels at each reference point approach target levels
according to figure 3-24. Fluctuations of lighting levels at reference points 1 to 4 seem to be
bigger than that of other simulations. Influences of the weighting factors might be the reason for
this. In terms of the mechanism of the factors, the system is assumed to find better solutions
which adapt to reference points with higher priority levels. In other words, in this model, lighting
levels at points 3 and 7 are expected to approach target levels most promptly. It is necessary to

49
focus on behaviours of each lighting level until the 150th generation, in order to investigate
detailed influences of the weighting factors due to the fact that fitness functions are decreased
dramatically until the generation.

Figure 3-25 shows differences of lighting levels of each reference point till the 150th generation.
For points 1 to 3, the weighting factor might work slightly to reflect user preferences. Lighting
level at point 3 which priority level is the highest approaches the target level most quickly and
that of point 1 might be the slowest. On the other hand, the factors do not influence the
behaviours of lighting levels of points 5 to 7. As figure 3-24 and 3-25 demonstrate, fluctuations of
lighting levels between reference points are not observed. Briefly, although reference point 7 is
the highest priority, the behaviour of its lighting level is the slowest in the same target level.

Thus, the weighting factors do not completely reflect user preferences in optimisation processes;
therefore, it is true that the weighting factor is not successfully achieved in the present version of
the RFDS. There might be several reasons for this, and further research is necessary to solve
this problem. However, the user preferences are considered to be essential to the design
process, so that future version of the RFDS will develop a better system of weighting factors.

Figure 3-23: Fitness scores of the simulation with weighting factors.

50
Figure 3-24: Natural lighting level of eight reference points.

51
Figure 3-25: Differences from each target level.

52
3.5.3 Benchmarks of the RFDS

In the previous sections, simulations are executed under mostly the same conditions including
population size, window size and so on. Meanwhile, it is also necessary for system
developments to measure benchmarks; therefore some models in different conditions are
simulated to compare the differences of responses.

Figure 3-26 explains the relations between population size and the initial fitness scores over the
calculation time of one generation. In this observation, population size is the only parametric
factor, and other elements such as room size, window size, and so on are the same as the
previous simulation models. According to the figure, calculation time for one generation is in
proportion to population size. In the case of 10 individuals, it takes around 200 seconds (3
minutes) in each generation, while calculation time for one generation increases up to 1200
seconds (20 minutes) if the population size is 90. In short, total calculation time until completion
might dramatically increase. As population size increases, the fitness score of the initial
population is improved.

Meanwhile, window size might be a more important aspect influencing the benchmark of the
system. Figure 3-27 demonstrates the relationship between window size and both calculation
time and chromosome length. In the simulation for this benchmark, window size is the only
parameter, and other elements are the same as the previous models. The results display an
asymptotic decrease, and the changes are more obvious in smaller window sizes, especially 100
x 100 and 200 x 200. The room size of the simulated model is very small; therefore total
calculation time will be much greater in practical building sizes.

If window size increases, total calculation time is decreased; however, the outlook of window
arrangements dramatically change as figure 3-28 demonstrates. Users might prefer smaller
window sizes in terms of aesthetic design aspects. According to this, the present version of the
RFDS might not be suitable for the design of larger scale of buildings. In terms of design ability
and comfort of the system operation, more research for quick responses or optimisation
processes are fundamental for further developments.

53
Figure 3-26: Initial fitness scores and calculation time in different population sizes.

Figure 3-27: Calculation time and chromosome length in different window sizes.

54
Figure 3-28: Generated windows in different window sizes.

55
3.5.4 Experimental Adaptation to Practical Design Process

In the previous three sections, several models are simulated to investigate the system
performances. According to the results and analyses derived from such experimental models,
the RFDS is applied to a more practical design process.

Figure 3-29 explains the simulation methods including arrangement of reference points, target
lighting levels, and so on. In this simulation, six small rooms are placed composing one building.
The size of each room is 3m cubed, and one reference point is placed in the centre of each room
at a height of 800mm from the floor level (desk height). The RFDS is expected to generate an
optimum façade design under these conditions. As far as the GA operators are concerned, each
parameter is initialised as follows: population size is 20, the number of generation is 350,
mutation rate is 0.03, and the number of elite is 2.

The optimised window arrangement generated by the RFDS is as figure 3-30 shows. The
number of generated windows on each wall is also demonstrated this figure. More windows are
arranged on the walls of rooms where reference points 2 and 3 with higher target level are
placed, than on that of rooms with lower target levels. Especially, there are few windows on the
walls of a room where reference point 4 is placed to avoid gathering much natural lighting.

Figure 3-29: Arrangement of rooms and reference points.

56
36 27
44
29
35

42

50
68

49 45

** Figures explain the number of


windows on each wall

Figure 3-30: Generated optimum window arrangement.

In this model, only one reference point was placed in each room so that the simulation process
was successfully and quickly achieved optimising room environments. Figure 3-31 and 3-32
demonstrate the generated solution and lighting analysis showing daylight levels of each
reference point. According to the figures, all the target levels are satisfactorily accomplished.
Additionally, lighting environments of each room are spatially optimised as well. In other words,
the system generates three types of internal environment related to target levels; the darker, the
normal, and the brighter room.

As far as the system performances are concerned, the RFDS works very well in this simulation
model. According to figure 3-33, the system is nearly converged around the 70th to 90th
generation in the condition that fitness function is almost 0. In addition to this, lighting levels of
each reference point also successfully and accurately approach target levels as figure 3-34
demonstrates. Unlike the previous simulation models, the system behaviours are exceedingly
effective. One of the reasons is the simple conditions required of the simulation as mentioned
above. Moreover, only one or two walls are used for the optimisation processes of each
reference point. Optimisation processes are much easier than with that of four walls like the
previous simulations.

Thus, the RFDS successfully generated the optimum façade under given conditions. Although
there are several problems in terms of further developments, it might be claimed that the RFDS
works well as a façade design tool.

57
Figure 3-31: Lighting analysis of the simulation model.

58
Figure 3-32: Lighting analysis showing lighting levels.

59
Figure 3-33: Fitness scores of the simulation model.

Figure 3-34: Natural lighting levels of six reference points.

60
Chapter 4: Future Developments and Possibilities of the RFDS

4.1 Other Search Methods

The GA used in the RFDS is a conventional genetic algorithm. In terms of calculation time, other
more advanced GAs or other search methods might be more suitable for the system. As far as
the present RFDS is concerned, according to the results of the simulations demonstrated in the
previous chapter, it takes at least one day to find solutions in terms of environmental
performance, although calculation time depends on the initialisation parameters. Needless to
say, the better qualities users require, the more time the system takes. There are no upper limits
for calculation times, and the total simulations can be above one week or more. Such heavy
processing loads are clearly not preferable situations for designers or architects.

GAs are blind search methods. It means that the algorithms are not suited to finding better
solutions in small ranges. That is to say, in the case that the algorithms are nearly converged,
fitter solutions are unlikely to be investigated. Moreover, premature convergences are also
possible for GAs. Due to the fact that accuracy is fundamental for environmental simulations and
such operations are delicate, better search methods need to be equipped in the RFDS.

There are various research studies about advanced GAs. Generally speaking, their
performances are developed in terms of processing loads and optimisation abilities. Therefore,
applications of such advanced technologies to the present RFDS is worthwhile for further
developments.

4.2 Additional Environmental Parameters and Multi-Objective Situations

Natural daylight level is the only parameter available in the present version of the RFDS. The
optimum solutions are generated under exceedingly simple environmental conditions; therefore,
it is not accurate and complex enough as a sustainable design tool. The present system needs to
equip more environmental parameters for advanced achievements.

There are several possible parameters for the RFDS, for example, internal temperature,
ventilation level, and so on. Internal temperature is one of the most important elements in terms
of environmental conditions as well as lighting level. Additionally, various benchmarks for
internal temperature are also available as well as lighting levels. According to this, internal
temperature is the most possible and necessary parameter for the next stage of the system. Not
only initial parameters are important factors for environmental optimisation, and fitness functions
are also deeply related to problem. As explained above, the system supports only daylight level

61
at present; therefore, the fitness function is also simply scored by only daylight level. In terms of
advanced environmental simulations, other fitness functions are required to be implemented.
There are various possibilities including, energy consumption, greenhouse gas emissions,
embodied energies, lifecycle costs, construction costs, and so on. Above all, energy
consumption or greenhouse gas emissions might be the most important elements due to the
present situation and social responsibility of the architectural field. As many people know,
architecture is the field in which energy consumption accounts for large part of the UK total.

Thus, more environmental parameters and other fitness functions are necessary for more
advanced systems in the future. Multi-objective situations are problematic aspects which cannot
be avoided in order to equip such additional environmental factors. In the case of fitness
functions related to a single objective, solutions are simply the maximum or the minimum;
however, in the case of multi-objective fitness functions, there might not be a 'best' solution. In
other words, as a certain aspect becomes better, other aspects have to become worse, and such
a fact is known as Pareto efficiency. Additionally, multi-objective optimisation problems are
related not only to parametric elements, but also to types of search methods; therefore, more
research is necessary to achieve developments of complex simulation systems.

4.3 Interactive Operations – Aesthetic Design and Environmental Optimisation

The main purpose of the RFDS is to design façades as well as to optimise window arrangements
in terms of environmental constraints. The design ability of the system is currently not high.
Several parameters such as window width, height, and interval, are available, and only one type
of window is generated by such parameters in one system operation. The RFDS does not
support a window increase operator, unlike the simple test model demonstrated in the previous
section, which generates various sizes of circular windows on walls. Furthermore, the
arrangements of generated windows completely depend on the environmental optimisation
process. Users are not allowed to control window arrangements in terms of aesthetic design.

As far as developments of design abilities are concerned, it is necessary to implement user


preferences as well as environmental optimisation. There are some emergent technologies to
achieve manipulation driven by human senses in the computational process. However, they are
still under development, and additionally, it might be difficult to manage both aspects. It is similar
to the situations of multi-objective optimisation problems.

62
4.4 Accessibility to the System

The present RFDS is a set of ECOTECT scripts which are a programming language to control
the software. Although the level of technical knowledge is required to run the script, only those
who have detailed knowledge of ECOTECT scripting are able to manipulate the current system
Because of this, the interface is one of the most significant aspects for system developments in
terms of accessibility to the RFDS.

The way to place reference points might need to be improved. Currently, users are allowed to
place several reference points in an ECOTECT scene; however, this way to determine target
points possibly causes complexity and unexpectedly heavy computational loads. In the case that
the density of reference points is too high, the RFDS might not be able to find good quality
optimised solutions. Even if the system could find good solutions, the calculation time becomes
surprisingly long because of longer generations or bigger population sizes. It might be difficult for
users to grasp the spatial distribution of target natural daylight level by using several points. One
of the possible improvements for this problem is a virtual distribution map of daylight level
generated by users. The map could be 2D grids or could be 3D cubes as Botsford used in the
Light House project. If the system could find solutions corresponding to the distribution maps, the
way to initialise target values is more user-friendly and it is easier to manipulate the internal
environment.

Meanwhile, initialisations of GA parameters are the most difficult and complicated factors for
ordinary users except for those who are good at programming. Generally speaking, designers or
architects are not experts with computers, although digital architecture is currently trendy and
fashionable. Therefore, proper initialisations of such parameters might not be expected. If users
are not able to utilise it accurately, the system is no longer effective, even if its simulation ability
is exceedingly high. According to such technical knowledge which is not directly related to
architecture, supportive systems to automatically or indirectly define GA parameters are
necessary. However, it is also important that convergences of the system are not deterministic,
and GA parameters are interactive in a complex way. Also, there must be a certain level of
technical knowledge expected from the users as this produces a technical output.

Thus, the RFDS is currently not suitable for all users, due to the complexity of parametric
initialisation and arrangement of reference points. However, the abilities of environmental
optimisations of the system are notable, and definitely effective for sustainable architecture
corresponding to energy efficiency, reduction of greenhouse gas emissions, and so on. If the
system becomes more feasible for every user and is widespread, the whole environmental
effects would be quite enormous. According to this, a user-friendly interface needs to be
developed.

63
4.5 Complex Geometry and Generative Design Tool

Due to the similar reasons to the accessibility to the system, the abilities of modelling need to be
improved in order to make it more widespread. Generally speaking, simulation software does not
support complex geometry, and as far as modelling is concerned, it is much less developed than
3D visualisation software, as the aims of these two systems are contrasting. Designers prefer 3D
visualisation software due to its flexible modelling abilities. It might be one of reasons why digital
technologies are not utilised as extensively in sustainable design as in aesthetic design. In other
words, in order to spread environmental simulation software and apply its abilities as much as
possible, the feasibilities of complex geometry and free form generation are fundamental.

Such a situation can be applied to the RFDS as well. One of possible solutions corresponding to
this problem is a combination of the RFDS and other platforms including CAD, 3D visualisation
and so on. If it is possible to transfer geometric data smoothly between two systems, the
advantages of both systems can be utilised. Complex objects are generated in other platforms,
and they are imported to the RFDS. According to this, high requirements toward modelling are
satisfied, and environmental simulations are effectively applied in the design process as well. In
addition to this, a form finding system like Genr8 could be developed by using such combination
technologies. If such possibilities are realised, boundaries between sustainable design and
aesthetic design do not exist any longer; furthermore, it encourages environmental simulation
technologies to be effectively applied to the design process.

64
Chapter 5: Conclusion

Annual energy consumption in the building sector is currently more than 45% of the UK total
energy consumption in all sectors. In terms of such a large ratio, research corresponding to
energy savings in the sector is a fundamental issue, and various aspects including insulations,
renewable energy sources, and so on, have already been investigated and improved.
Computational optimisation of the natural environment for sustainable design outcomes lags
behind its use in other related fields. In this thesis, the novel sustainable design methods
combined with emergent digital technologies were researched. Case studies of several
completed projects related to digital sustainable design demonstrated present situations,
potentials and efficiencies, and experimental models developed by the author found practical
possibilities and future developments of the combination of sustainable design and digital
design.

Digital technologies, especially evolutionary computations including genetic algorithms, can be


regarded as a useful and effective method according to the results of the case studies in chapter
2. One of their remarkable abilities of these technologies is the processing speed to calculate,
and simulation skills are far better than that of the human brain. In fact, such technologies
utilised in some projects to optimise energy consumption or natural environment produced better
solutions than those generated by humans.

However, although there are advantages produced by the use of emergent technologies for
sustainable design, several problematic aspects become clear. Multi-objective optimisation
problems in relation to Pareto efficient situations might be the most significant issue. Sustainable
design is more complicated than some other architectural design fields due to the fact that
complexity including the natural environment, materiality and so on is fundamental for the design
process. Designers are required to consider overall aspects. For instance, in the aesthetic
design, “beauty” is the only important factor; however, it cannot be applied to sustainable design.
Even if one element is optimised, it does not demonstrate overall optimisation, and on the
contrary, unexpected bad effects in other elements can be found. Because of this, it is difficult to
determine what the 'best' solution is in many situations.

Relationship between the virtual and the practical are the most problematic. In short, it is how
computers recognise practicalities in a virtual space. Objects on the screens of our computers
are sets of digital data which are binary codes; simply, “1” and “0”. Therefore, although objects
on the screen are different from each other for us, every element is the same for computers. As
demonstrated above, accuracy is very important, and such computational problems are
obstacles for digital sustainable design.

65
In chapter 3 and 4, some aspects related to system popularisation are discussed in terms of
analytical points derived from the experimental models. It can be stated that although they are
suitable for sustainable design, not many people are familiar with digital technologies at present.
Accessibility to them is one of the most important elements stopping digital technologies from
becoming more widespread. The experimental model demonstrated the complexities of
operations of the system for instance, parameter initialisations, and their need to be operated
accurately in order to fully utilise the system. Such technical requirements currently
uncomfortably straddle the architectural and the science field. The majority of designers or
architects do not have any knowledge of digital technologies and do not handle them. Therefore,
in terms of popularisation of powerful digital tools, easy access or manipulation of the systems
needs to be solved.

The experimental models also demonstrated that one of the reasons why simulative methods
are generally not popular as design tools might be the low modelling ability of the simulation
software. Such software generally does not support free form generation or complex geometry.
However, “beauty” is exceedingly important in the architectural fields, for designers. According to
this, simulation tools are not useful for them. They might prefer to use high-end 3D visualisation
or drafting software in their design process for flexible modelling ability. In addition to this, the
difficulties of data transfer between different systems might also hold back the design process.
Platforms used by architects are generally CAD or 3D tools; however, simulation software does
not support the file format used by such platforms. Therefore, it is difficult to choose systems
according to purposes. Thus, the boundary between simulation tools and others exists as an
obstruction to an effective combination of sustainable design and digital technologies.

As stated above, emergent digital technologies have potential and are useful; meanwhile, they
are still under development, and there are various problems to be solved in order to fully apply
them to sustainable design. However, according to applications in other fields and their
capability, digital sustainable design in generative and responsive optimisation process can also
become a new method resulting in enormous environmental influences including energy
consumption, greenhouse gas emissions, and so on. In addition to this, some complete projects
and the experimental model in this thesis also indicated their potential in environmental design.

66
Bibliography

Andrasek A. (2006). All-Over, Over-All: biothing And Emergent Composition. Architectural


Design 76 (4), pp.72-81.

Aranda/Lasch (2006). Tooling. New York: Princeton Architectural Press.

Baglione, C. (2006). Casa a Notting Hill. CASABELLA 744, pp.4-11.

Ball P. (1999). The Self-made Tapestry: Pattern Formation In Nature. USA: Oxford University
Press

Caldas, L. (2001). An Evolution-Based Generative Design System: Using Adaptation to Shape


Architectural Form. Unpublished Doctoral thesis, Massachusetts Institute of Technology, USA.

Caldas, L. and Norford, J. (2001). Architectural Constraints in a Generative Design System:


Integrating Energy Levels. Proceedings of Building Simulation 2001, August 13-15. Rio de
Janeiro.

Caldas, L. and Rocha, R. (2001). A Generative Design System Applied to Siza’s School of
Architecture at Oporto. Proceedings of CAADRIA ’01, Sydney, Australia, 19-21 April.
pp.253-264.

Caldas, L., Norford, L. and Rocha, J. (2003). An Evolutionary Model for Sustainable Design.
Management of Environmental Quality: An International Journal 14(3), pp.383-397.

Frazer, J. (1995). An Evolutionary Architecture. London: AA Publications.

Frazer, J. and J. (1998). The Groningen Experiment: Architecture as an Artificial Life Form –
Materialization Phase II. Architectural Design 68 (11/12), pp.12-15.

Frazer, J. H., Frazer, J. M., Xiyu, L., Tang, M. X., and Janssen, P. (2002). Generative and
Evolutionary Techniques for Building Envelope Design. In: Generative Art 2002, 5th International
Conference GA2002, Milan: Politecnico di Milano University.

Gianni Botsford Architects (1996) Data Structure Research. [Online]. Retrieved on 16 January
2008 from http://www.giannibotsford.com/projects/012/index.php

Gianni Botsford Architects (1996) The Groningen Experiment. [Online]. Retrieved on 16 January
2008 from http://www.giannibotsford.com/projects/019/index.php

Goldberg, D. (1989). Genetic Algorithms in Search, Optimization, and Machine Learning.


Canada: Addison Wesley.

67
Hagan, S. (2008). Digitalia: Architecture and the Digital, the Environmental and the Avant-Garde.
London: Routledge.

Hemberg, M. (2001). GENR8 – A Design Tool for Surface Design. Unpublished Master thesis.
Chalmers University of Technology.

Hemberg, M. and O’Reilly, U.M. (2004). Extending Grammatical Evolution to Evolve Digital
Surface with Genr8. In: Genetic Programming 7th European Conference, EuroGP 2004,
Proceedings. Berlin: Springer-Verlag.

Hemberg, M., O’Reilly, U.M., Menges, A., Jonas, K., Goncalves, M. and Fuchs, S. (2006)
Exploring Generative Growth and Evolutionary Computation for Architectural Design. Chapter
submission to Art of Artificial Evolution.

Hensel M. and Menges A. (2006). Morpho-ecologies. London: AA Publications.

Hensel M. and Menges A. (2006). Differentiation and Performance: Multi-Performance


Architectures And Modulated Environments. Architectural Design 76 (2), pp.60-69.

Luebkeman, C. and Shea, K. (2005). CDO: Computational design + optimisation in building


practice. The Arup Journal 40 (3), pp.17-21.

Makoto sei Watanabe/ Architects’ office (1990) The Induction Cities. [Online]. Retrieved on 15
January 2008 from http://www.makoto-architect.com/idc2000/index/html

Makoto sei Watanabe/ Architects’ office (2000) Subway Station/ Iidabashi. [Online]. Retrieved on
15 January 2008 from http://www.makoto-architect.com/subway/syb_e.htm

Makoto sei Watanabe/ Architects’ office (2004) Tsukuba Express / Kashiwanoha-Campus


Station. [Online]. Retrieved on 15 January 2008 from
http://www.makoto-architect.com/three_st/tx_01.htm

Makoto sei Watanabe/ Architects’ office (2005) ShinMinamata MON. [Online]. Retrieved on 15
January 2008 from http://www.makoto-architect.com/minamatamon/smm01.htm

Makoto sei Watanabe/ Architects’ office (2006) KeiRiki-2 : INDUCTION DESIGN-V. [Online].
Retrieved on 15 January 2008 from http://www.makoto-architect.com/KEIRIKI-1/keiriki.htm

Menges A. (2006). Self-Organisation and Material Constructions. Architectural Design 76 (2),


pp.34-41.

Mitchell, M. (1998). An Introduction to Genetic Algorithms. Cambridge: MIT Press.

68
O’Reilly, U.M., Hemberg, M. and Menges, A. (2004). Evolutionary Computation and Artificial Life
in Architecture: Exploring the Potential of Generative and Genetic Algorithms as Operative
Design Tools. Architectural Design. 74 (3), pp.48-53.

O’Reilly, U.M. and Hemberg, M. (2007). Integrating Generative Growth and Evolutionary
Computation for Form Exploration. Genetic Programming and Evolvable Machines. 8 (2),
pp.163–186.

Rahim A. (2006). Catalytic Formations: Architecture and Digital Design. Abingdon: Taylor &
Francis.

Sims, K. (1994). Evolving 3D Morphology and Behavior by Competition. In: Brooks, R. and Maes,
P., (ed.) Artificial Life IV Proceedings of the Fourth International Workshop on the Synthesis and
Simulation of Living Systems (Complex Adaptive Systems). Cambridge: MIT Press, pp. 28-39.

Sims, K. (1994). Evolving Virtual Creatures. In: Computer Graphics: Proceedings : Siggraph 94
Conference Proceedings July 24-29, 1994, Association for Computing Machinery.

Terzidis K. (2006). Algorithmic Architecture. Oxford: Architectural Press.

Xiyu, L., Frazer, J., and Tang, M. X. (2002). A generative Design System Based on Evolutionary
and Mathematical Functions. In: Generative Art 2002, 5th International Conference GA2002,
Milan: Politecnico di Milano University.

Xiyu, L., Mingxi, T. and Frazer, J. (2005). An eco-conscious housing design model based on
co-evolution. Advances in Engineering Software 36, pp.115–125.

69
Appendix – A

This appendix is the reference manual of the RFDS

70
71
72
73
74
75
76
77

S-ar putea să vă placă și