From Optimisation to Generation: Evolutionary processes in architectural design

Acknowledgement

I would like to thank Emmanouil Zaroukas for his support with the research and comments on the draughts of this paper.

Table of content

Introduction

A computer is, essentially, the same as a huge army of clerks, equipped with rule books, pencil, and papers, all stupid without initiative, but able to follow exactly millions of precisely defined operations… In asking how computers might be applied to architectural design, we must therefore ask ourselves what problems we know of in design that could be solved by such army of clerks…At the moment there are very few such problems. (Alexander, 1967, pp. 8-12)

Computers certainly made their way in our everyday lives. They are such versatile machines it is no surprise they are extensively used for work, communication and even leisure. In the area of architecture, computers were first introduced as replacement for traditional drawing boards. For this reason, they were quickly adapted to fit the preestablished methods of design in means of manual drafting of production drawings and geometry. Ever since architects saw a glimpse of computers’ potential, a vast number of computer-aided design (CAD) tools have been and currently are being developed. Such continuous improvement of technology could reasonably make one wonder what the full potential of computer in design could be. The main challenge lies in the cultivation of computation as a tool that complements the designer’s capabilities in the conceptualization and production of design artefacts in the contemporary architectural agenda (Ahlquist & Menges, 2011). Therefore, to investigate their potential within a human-centred design process, a symbiosis of architects and computers, we should step back and revisit the use of computers in architectural design. Do they have capacity to perform design tasks without being explicitly programmed how to do them, but instead what is to be done?

Similar question has already been asked by Holland (1992, p. 1) in the beginning of his book Adaptation in natural and artificial systems. Holland turned to biology to answer his question and later developed genetic algorithms, which are now widely used in various fields. Biology and architecture have a lot in common, too. By studying of the modern synthesis of evolutionary biology and understanding its mechanics one could draw parallels between the roles of reproductive systems and architects, DNA molecules and production drawings, parents and actual built buildings. There are many biological analogies in architecture and applied arts in the work of Steadman (2008), complemented by referral to nature as a blind designer by Dawkins (2006, p. 50), and the comparison of cities to living organisms  (West, 2017, p. 321) and list goes on. These show strong ties between biology and architecture and give good predisposition to suspect the adoption of nature’s way to design may have practical application in architecture. For this reason, the paper shall investigate further the use evolutionary processes in architectural design. Despite the distinct nature of evolutionary design methods, the circumstantial evidence gives reason to speculate that evolutionary processes can: 1. allow computers to automate and speed up typical design tasks, 2. be utilised to optimise design artefacts and 3. be used for generation of novel solutions.

The foundations of the subject have already been laid and the paper will aim to present the past achievements on the topic thought critical analysis of texts and case studies. Based on the above evidence, then the effective use of evolutionary design methods will be critically assessed in regard to architectural design in practice.

Section I

Literature review and background information

In biology the term evolution explains the gradual change of organisms. There have been number of evolutionary theories, the most popular of which are perhaps the Darwinian and Lamarckian evolution. In the mid nineteenth century Charles Darwin published his book The Origin of Species by Means of Natural Selection, in which he described his observation evolution of biological organisms. He established a new world which broke away from the Newtonian paradigm of stability – a world in a continuous process of evolution and change. (Frazer, 1995, p. 13) Darwin came to three rules that drive the evolutionary process by observing populations of finches on the Galapagos islands, which let him formulate his theory:

  1. Living creatures reproduce at much higher rate to simply maintain their numbers. Therefore, populations generally become larger.
  2. Even though dramatic changes certainly do occur in nature, the general number of members in a population from any given species remain relatively unchanged from one generation to another. The two observations suggest that a struggle for survival must take place.
  3. All members within a species are not identical. There was variation of inheritable traits. (Steadman, 2008, p. 73)

Darwinism is a theory of biological evolution based upon inherited variations in organisms and natural selection of fitter variants to produce species adapted to their habitats. It could be described as a process of gradual change of organisms due to non-random selection of random variations and nature being the driving force by acting as selection criteria. The variation of traits in organism of the same population is due to random errors, or biologically speaking mutations, which could only happen during the process of reproduction. These mutations later affect the development of the organism. If the changes caused by a given mutation happen to be beneficial and improve the short-term survival, or fitness, of the organism then they are likely to reproduce and pass them to the next generation. Then again, if the changes happen to worsen organism’s survival then they are likely to be lost should the organism die and fail to reproduce. Given enough time, this process of passing characteristics from parent to offspring results in gradual improvement, known as evolution.

On the other hand, one of Darwin’s predecessors, Jean-Baptiste Lamarck (1916), had different theory of evolution, which he based on heredity of acquired characteristics. Contrary to changes occurring only during reproduction process as suggested by Darwin, Lamarck suggested that it is possible for characteristics acquired during organism lifetime to be passed from parent to offspring. The more famous examples given by this theory are the giraffe’s neck being a result of the parent stretching through his life and passing the same characteristics to the offspring or a blacksmith’s son’s muscles being passed by his father. Lamarck’s theory is generally discredited in relation to organic evolution as it is not possible for such bodily changes acquired during one’s lifetime to affect the genetic material directly and be passed to future generations. Even though his theory is not valid in respect to physical characteristic, it turned out to be true for culturally acquired characters or simply passing mental concepts from generations. (Steadman, 2008, p. 124)

Going back to the blacksmith’s son example, even though his father cannot pass the muscles’ strength, he certainly can pass his knowledge and experience. More so, Steadman sees tools as organs or as extensions of the physical body and goes on saying:

There is thus a very metaphorical sense in which one might see man and his material creations together as some kind of hybrid mechanical/organic creature, in which the processes of evolution go on at some speed in the mechanical parts by comparison with very much slower changes in the organic parts. From this point of view cultural evolution and specifically technological evolution is seen as a continuing phase of biological evolution in man, proceeding by different mechanisms, and overlaid onto the Darwinian, genetic process. (Steadman, 2008, p. 120)

For the later research in this paper the term evolution will be used and in its Darwinian definition and will be elaborated further to explain the mechanics of biological evolution in respect to more recent advancements in genetics. However, it should be noted that one could reasonably see the later deployment of designer’s knowledge to a computer as the technological evolution Steadman is referring to. Designers and programmers have a lot in common, thus when discussing the computational side of evolutionary design the paper shall address the designers as programmers, however the difference between both is rather minimal.

Going slightly back to the opening quote to remind the reader, Alexander (1967, pp. 8-12) asks what problems we know of in design that could be solved by computers. There are in fact a number of approaches to architectural design practised by architects. This is why it is not easy to define and formulise such design problems in architecture. For Alexander himself design could be described a as problem of optimisation, similar to the gradual improvement of organism in nature. In his Notes on the Synthesis of Form he states:

 “[…] Every design problem begins with an effort to achieve fitness between two entities: the form in question and its context. The form is the solution to the problem; the context defines the problem.”  (Alexander, 1964, p. 15)

The author defines form and context as creating an ensemble, with a good fit being a desired property. The terms fit and fitness are commonly used by Darwin and Dawkins in their work on biological evolution and this seems to be acknowledged by Alexander describing biological ensemble as being made up of a biological organism and its physical environment.

In a way, Frazer (1995) not only takes Alexander’s idea of an ensemble but adds on to it by defining architecture as a metabolism or an adaptive system. He pursues novelty in architectural design by investigating form-generating processes. Generally speaking, Frazer describes architecture is a living organism, capable of biological growth and response to unpredictable environments. For this reason, the paper shall address the use of evolutionary processes in architectural design as problem of optimisation and generation, seeking of novel solutions.

Frazer (1995, p. 60) also critiques the trend in computer-aided design (CAD) tools reflecting traditional non-CAD design methodologies. As known, since the 1960s the area of computation has developed rapidly and made solid presence in the field of architectural design. Various design and CAD methods have been developed. Gero (1994) presents two main categories CAD tools development. Traditional CAD tools coming of-the-shelf, mimicking manual drafting and modelling and analysis software usually fall under the first category. They are often used to improve efficiency and automate simple tasks. The second category consists of knowledge-based tool to carry the synthesis of designs. However, Frazer (1995, p. 60) claims that present architectural design processes are fundamentally wrong and not worth imitating. In contrast, the second category which Gero talks about gives birth to novel generative approach to design exploration and it happens to be the main area of interest.

Problem and hypothesis

Architectural design processes and evolution in nature have a lot in common. They both use their available resources to produce design artefacts or biological organisms respectively. Some architects tend to adopt exemplars by borrowing the best aspects of existing buildings and combining them in what they believe is the best way to produce new designs. Similarly, in nature organisms reproduce. In Darwinian evolution, the current generation pass their inherited characterising to the future generations and so on. In this case, the original building exemplars could be seen as parents to new generation of architectural design artefacts. The results of both processes are the many generic building types in architecture and variation of species in nature. This manual trial and error method however does not allow the exploration of variety of design artefacts as it takes to long for architects to evaluate them against performance criteria. Nature is by far the most complex system known, and yet all biological organism seems to be perfectly adapted to their natural environment, creating a large eco system with vast number of species. This is due to biological evolution process gradually changing organisms to better fit their environment through on-going optimisation by natural selection. In this context, evolutionary design involves borrowing ideas from natural evolution, combines them with CAD and analysis software and it is then being used in design for optimisation and generation for the purpose of the paper. (Bentley, 1999, p. 35) The term evolutionary design in the following parts of the paper could be best described as the activity of design itself made into a Darwinian process. (Steadman, 2008, p. 248) One could see origin of evolutionary design as being quite distinct as it can be traced to the intersection of computer science, design and evolutionary biology (fig.1).

Figure 1: Roots of evolutionary design

As noted above, evolution generally increases the fitness of organisms and making them better adapted to their habitat. On one hand, evolutionary approach could potentially solve design optimisation problems in architecture like Alexander’s endeavours to achieve good fit between form and context. This could happen by replicating digitally the evolutionary processes in nature and harvesting their optimisation power towards goal-oriented architectural design problems. On the other hand, evolutionary design could bring novelty for architects seeking for tools for generation. This could be achieved by both broadening architects’ solution space for design exploration and providing new ways to navigate to through this space. Notably, these proposals will not be possible by using computers as overcomplicated drawing boards and without them having the capacity to simulate evolutionary tasks. Despite the distinct nature of evolutionary design, the presented evidence gives reason to speculate upon its possible application in architectural design. Use of evolutionary design methods shall then be assessed before concluding whether they should be introduced to architects’ design repertoire. All things considered, the intentions of the following research are not to exclude other possible uses of evolutionary processes to architectural design, nor to undermine the use other CAD methods. The given propositions would not be possible prior to the advancements in biology and capacity of computers to be programmed to perform such evolutionary tasks. These advancements will be presented and explored further below.

Section II

Borrowing from nature

It is important to point out the dual meaning of the word design before proceeding further with the enquiry. Until now it was used to define both the activity of designing an object (design as act) and design as product of this act (as an artefact or object). For the rest of the paper design will be used as to the activity of designing, while in the occasions when used to describe an object it will be referred to as design artefact or solution.

Nature has frequently been used as a source of inspiration for architect. It is little to no surprise to see natural processes being currently used in design. Understanding the mechanics of those processes of biological context should allow one to better comprehend the principles behind the evolutionary design. Darwin, despite the great discoveries, had limited information on how the variation in populations actually occurs and how characteristics were passed from one generation to another. The popular theory at the time to explain the passing of characteristics from parents to offspring was the blending inheritance theory, meaning traits from parent organisms were averaged before being passed to future generation. This blend would eventually result for almost any variation to be lost. And it wasn’t until the late 19th century when Gregor Mendel, known as father of modern genetics, explored the idea of heredity units, now called genes, which replaced the theory of blending inheritance. The combination of Darwin’s observations on survival of the fittest and Mendelian inheritance in early 20th century was to be known as the modern synthesis of evolutionary biology or Neo-Darwinism. (Dawkins, 2006, p. 113)

More recently Dawkins (2006) took the Neo-Darwinian idea and elaborates further on the its mechanics. According to Dawkins (2006, p. 112) living things are collections of molecules, like everything else. The unique property of living organisms’ molecules is that they can be put together in much more complicated patterns than the molecules of inorganic things. This construction of complicated patterns is done following assembly instructions inside organisms themselves. It is the assembly instructions that are of great importance. For organic beings this set of instructions is embedded in the DNA molecules. The DNA hold the all genetic material (genome) or all instructions needed for the growth, development, functioning and reproduction of organisms. All the DNA in organism has permanent address and content (Dawkins, 2006, p. 117). This means that there is a specific location in organisms’ DNA which could be, and in fact is, revisited for its content to be read multiples times. Although, the content cannot be altered during individuals’ lifetime, leaving behind Lamarck’s theory. DNA is arranged in along strings called chromosomes (Dawkins, 2006, p. 117). People for example have 46 chromosomes, while other animals have different number. In human reproduction when a new organism is being produced, it takes 23 chromosomes from each parent and they are grouped in pair to make to total number of 46. The genetic material of the parents could be reshuffled and combined in number of ways, but it is only during the process of conception that the unique genetic code of individual organism is generated. Occasionally random errors (mutations) do occur in nature. DNA has a self-replicating property and mutations are precisely errors in this self-replication process. (Dawkins, 2006, p. 129) If an error happens during the self-replicating process of DNA it cannot be undone or altered, but it is possible for this new mutated genetic code to be passed to the offspring during reproduction and as seen above. This mutation then affects the growth and development of the new organism and has a chance to be passed to future generation if successfully overcomes the natural selection.

In his book Dawkins (2006, p. 46) gives example of a biologically inspired computer program to demonstrate the difference between single-step and cumulative, incremental step-by-step, selection. The program begins by generating a sequence of 28 random characters, the amount of characters is the phrase – METHINKS_IT_IS_LIKE_A_WEASEL, which acts a target phrase for evaluation. Then, the program replicates the original phrase, but it is given chance of a minor random error happening, thus giving a variation similar to nature to occur. The computer then evaluates the new mutant phrase for resemblance to the target phrase. In other words, it evaluates the fitness of the phrase to the environment. If the mutant phase happens to be closer to the target, no matter how slight the improvement is, it is then used as a parent for the next generation. The program loop of replication of the current best solution (parent) phrase and evaluation runs until the target is reached. In the first run of the program Dawkins (2006, p. 43) reported that the target phrase METHINKS_IT_IS_LIKE_A_WEASEL was achieved in 43 generation, while in the second run the number was 60 generations and in the third 41. This is a significant improvement over the small odds of landing on the target phase in a single-step thus it validates the optimisation properties of evolution by cumulative selection. The described program Dawkins created was a very brief example of evolutionary computation, but it captures the general use of evolutionary processes towards a specific goal. Of course, nature does not have a long-term goal, but instead improves the short-term survival of organisms. This is why Dawkins (2006, p. 50) decides to refer to nature as the blind watchmaker. If agreeing with the above, one could reasonably conclude that each organism within a population carries unique genetic information, ready to be transferred to future generation as long as nature criteria is satisfied. It is important at his point to note the genes themselves are not directly evaluated, but their manifestation in organism’s body. As mentioned earlier genes are a set of instruction embedded permanently in individual’s DNA. These instructions are then executed during the growth and development of an organism. The later physical characteristics and attributes resulted by following these instructions are known as phenotype, while the specific genes in individual’s genetic make-up are known as genotype. Nature evaluates the complete organisms, not the assembly instructions. Thus, individual’s genotype is only indirectly selected, but the phenotype is the one fighting to survival.

This chapter presented how the Darwinian evolution was combined with the Mendelian inheritance to establish the modern synthesis of evolutionary biology. This greater knowledge of genetic transmission enabled Dawkins to write a computer program and recreate virtually the natural evolutionary processes. The actual making of such evolutionary programs and algorithms shall now be detailed and explained.

Evolutionary Computation

As it can be seen above, there have been remarkable discoveries recently in the area of genetics that gave almost new dimension for understanding how organisms evolve. And yet there are undoubtedly more discoveries to be made in future. For this reason, it is not possible to explicitly program computers to evolve solution as evolution is not yet fully understood. (Bentley, 1999, p. 5) In reality, evolutionary algorithms or EA are being written to closely replicate natural processes and evolutionary improvement of artefacts in due to the optimisation property of the algorithms. (Bentley, 1999, p. 5)  There are four common types of EAs developed independently since the 1960s and still used today. (Bentley, 1999, p. 7)  Those are: genetic algorithms developed by John Holland and later made widely known by David Goldberg, evolutionary programming  created by Lawrance Fogel and his son David Fogel, evolution strategies  by Ingo Rechenberg and Thomas Back and lastly genetic programming – John Koza.  There are some differences between the different algorithms, but Bentley (1999, p. 25) argues the general concept behind them is hardly different and then presents a general architecture of evolutionary algorithms or GAEA (fig.2).

Figure 2: General architecture of evolutionary algorithms (GAEA)

By using the available information from evolutionary biology Bentley (1999, p. 6) describes how an EA could be programmed. He begins by saying the computer is to be instructed to have number of solutions, thus to maintain a population, a rule coming from original Darwinian theory. In reality nature is the selection criteria. In EA however, solutions they are given fitness values. (Bentley, 1999, p. 7) These are typically positive real numbers with perfect fitness score being a zero. Each value is determined according to a fitness function given by the programmer, the resulted values are then evaluated to global selection criteria. In every population the better performing solutions are allowed to reproduce and worse performing ones to die. Identical to nature, each new generation of solutions inherit they parent characteristics with some minor variation (mutation) occurring. The process then goes on to minimise solutions’ fitness values until it they ultimately reach zero. Expectedly after number of cycles the evolved population of solutions become seemingly better than original population as similarly seen earlier in Dawkins experiment.

A computer can independently process the input data. At first look, one may conclude that a computer is given certain level of autonomy by its maker by executing an EA. Although this may be true, the programmer still has a great control over each stage of presented in general architecture of evolutionary algorithms. Bentley (1999, p. 430) defines individuals as the collection of data that makes up a member in a population, its genotype, phenotype and fitness values, but more aspects could be added if required. As known, the genotype is the genetic make-up of individuals and it is first to be represented as numeric values.  As mentioned before, in biological organisms the genetic information in the DNA is addressed, in this case the programmer can define the number of addresses and their content in variables. The collection of all genotypes defines a search space. (Bentley, 1999, p. 51)  Therefore, one could see each gene as a new dimension given of the search space to be explored (fig.3). This is in fact a crucial step because the more genes are present in a chromosome the greater the search space to be explored. If the search space is too great this could result in the process of evolution to slows down or even to stop entirely. (Steadman, 2008, p. 254) Phenotype, as outlined, is the manifestation of the genes in a member, the solutions to be evaluated, and the collection of all the phenotypes it defines a solution space. The difference made between search space and solution space, between genotypes and phenotypes in only made in GA while the other three approaches, EP, ES and GP, make no such distinction. (Steadman, 2008, p. 251)

According to Bentley (1999, p. 5) evolutionary computation is all about search. In computer science, search algorithms define a computational problem in terms of search, where the search-space is a space filled with all possible solutions to the problem, and a point in that space defines a solution (Kanal & Kumar, 1988 cited in Bentley, 1999, p. 5) . Computers can continuously evolve solutions due to EA having feedback loop. Once computer finds a population of solutions by going through all necessary stages listed in Bentley’s GAEA, it continues on searching generation after generation.

The entire evolutionary process taken by the machine is could be defines as rather unconscious and the generated solutions arise after gradual mindless improvements. (Bentley, 1999, p. 2) There are also other search algorithms apart from evolutionary ones such as hill climbing, simulated annealing and more, but these fall outside the scope of the enquiry. (Bentley, 1999, p. 4) Despite the powerful properties of EA to search and evolve solutions, it would seem naïve to expect practical solution should a programmer neglects his role in the evolutionary process.

Figure 3: Genotype to phenotype mapping

For this reason, it could be said the successful use of any EA is highly depended on the perceptual and cognitive abilities of a programmer.

So far, the paper presented that by using EA computers are able to establish a certain level of autonomy if problem could be defined in evolutionary terms. This not to say the intentions are to look at computers as fully independent designer, but to demonstrate they can be given freedom to process and interpret data in order to evolve solutions. In fact, the relationship between the programmer and the computer could be described as rather symbiotic. Just as Steadman (2008, p. 120) suggests looking at tools as extension of the humans’ physical bodies, by externalising human logic into algorithm a programmer can extend his cognitive abilities and increase his design potential. Thus, algorithms could be seen as extensions of the human brain. (Terzidis, 2011, p. 94) The first point raised in the beginning of the paper was that computers can automate tasks and produce solutions, without being given explicit instruction on how to do it. As presented, if a problem is could be encoded in evolutionary terms a computer can, to a certain degree, independently process information and evolve solutions. By writing and using evolutionary algorithms and deploying design knowledge into these, one can harvest the available computing power to automate the production of improved solutions to predefined criteria. This could then lead to the displacement of designer’s intelligence from manual production of artefacts to presenting a given problem as Darwinian process to be solved computationally.

Section III

Evolutionary algorithms in architectural design

Coming back to architecture, the paper took two design problems – optimisation and generation. They seem completely contradictory. The first one is looking to find the single best solution to a well-defined criterion, for example a building’s orientation for maximum sunlight gain. Whereas the problem of generation looks for ways to produce multiple and novel solutions, therefore to give architects the opportunity to explore variation on possible design artefacts within a given set of requirements. However, if both problems are to be defined in terms of computation terms, we see something peculiar. Both of them could solved by search and it is fact what architects do when designing building – trying the produce the more satisfactory design artefact to meet the requirements they have.

According to Bentley’s GAEA evolutionary algorithms work in a similar way, however he defines four aspects of evolutionary design: Evolutionary Design Optimisation, Creative Evolutionary Design, Evolutionary Art and Evolutionary Artificial Life Forms. (Bentley, 1999, p. 36) As seen, the native function of evolutionary algorithms is to evolve solutions by improving and optimising their fitness. For this reason, it seems appropriate to begin the investigation of how evolutionary processes are used in architectural design with optimisation orientation.

Problem of Optimisation

At the time of their publication,  Alexander’s Notes on Synthesis of Form have been positively accepted within the architectural community and presented great potential of becoming a contemporary architectural theory. (Steadman, 2008, p. 163) Alexander begins his Notes by giving simple example with the designer decision underlying the material choice for a mass production item such as vacuum cleaner. For this example, Alexander (1964, p. 3) states that there are contradicting aspects of the vacuum cleaner design like simplicity, performance, jointing and economy (cost). Assuming the desire of the designer is to minimise the cost by choosing the most cost-effective material for each component of the vacuum cleaner, this will most likely affect negatively the other aspects. Whereas, greater variety of materials may over-complicate the jointing and simplicity, as well as reduce the performance. On the other hand, if a designer would choose the materials to result increased performance of the vacuum cleaner, this will likely increase the cost of the product and not necessarily benefit the simplicity and jointing. Alexander (1964, p. 3) explains that this is a typical design problem. Here is probably a good time to remind the reader about the increased quantity of information and requirements introduced recently to architecture in historic plan such as sustainability, acoustics, new materials and construction methods and more. All of them existing simultaneously and interacting with each other, thus raising the difficulty of finding appropriate design solution. Defined in this way, adding any more requirements would eventually result an exponential increase of interaction links to be analysed one by one. Similar problems are not foreign in computing. The exponential increase of possible scenarios to be evaluated is commonly combinatoric explosion. This is a fundamental problem in computing. One of the most famous examples is perhaps the travelling salesman problem. Virtually speaking there are no (defined) limits to man’s cognitive capacity to multitask, however practice shows there are limitations of some kind varying between individuals.  (Alexander, 1964, p. 5) As a consequence of these limitations, complexity makes the problem of discovery of optima a long, perhaps never-to-be-completed task, and only the best sound options must be explored at a time. (Holland, 1992, p. 1) As many probably suspect, Alexander presents more of analytical approach to problem solving underpinned with mathematical logic, which might appeal more to certain types of architects and designer and be criticized by others.

The adaptation of biological evolution in architectural design is not extensively discussed in his Notes. However, he states that “[…] Every design problem begins with an effort to achieve fitness between two entities: the form in question and its context. The form is the solution to the problem; the context defines the problem.” (Alexander, 1964, p. 15). By doing so, the author presents form and context as creating an ensemble, with a good fit being a desired property. The terms fit and fitness are commonly used by Darwin and Dawkins in their work on biological evolution and this is acknowledged by Alexander describing biological ensemble being made up of a natural organism and its physical environment. However, the biological analogy is not developed further in the book. Instead, the proposed form-making process is being describes as: “a negative process of neutralising the incongruities, or irritants, or forces, which cause misfit.” (Alexander, 1964, p. 24). The term ensemble is later replaced by system relaying on inner organisation. (Alexander, 1964, p. 18). An analogy of this adaptation through refinement is given with finding equilibrium in cybernetics. (Ashby, 1954 cited in Alexander, 1964, p. 196) For the purpose of the essay the qualitative judgement of ‘good fit’ is left to individual designers as aesthetics in design is a deep topic on its own and the focus of the paper will remain process of finding a ‘good fit’.

As a solution to his enquiry, Alexander shows an example how one can create and use independent abstract patterns, or diagrams, to map and resolve a small system of interacting and conflicting forces. Due to the diagrams being independent he suggests that they could be improved one at a time and then merged and combined in many ways to create not only one, but variety of designs.    After the notes were originally published, he dedicates significant time on developing his method and discovered the diagrams themselves had immense powers. However, 10 years later, in the preface to the paperback edition of the book Alexander (1964) admitted the process he previously described for creating these diagrams was unnecessary due to being too complicated and formal. Steadman (2008, p. 194) later argues that Alexander’s did not succeed in applying his proposed methods in practice is due to biological fallacies. By diminishing the role of the architect, Alexander expects the contextual forces to almost create form by themselves. (Steadman, 2008, p. 194) Furthermore, Steadman suggest that Alexander’s research may have led him to propose a kind of simulated, and hence speeded up, version of technical evolution, carried on in the drawing office or in the ‘design laboratory’, and using mathematical or computer models to represent form, context and their interaction. (Steadman, 2008, p. 195 & 249)

All things considered, an optimal solution to satisfy number of requirements as described by Alexander as typical design problems and perhaps similarly seen by many other architects seems as an adequate task to be given to an evolutionary algorithm. (Steadman, 2008, p. 195 & 249 and Frazer, 1995, p. 17)

As discussed before, using CAD to mimic traditional drafting methods would essentially mean typical trial and error process of designing by gradual improvements until a design artefact achieves meet the given requirements, but the process in lengthy and does not allow many options to be explored. In order to address this issue, the paper shall look at the first of the four aspects presented by Bentley (1999, p. 35) – Evolutionary Design Optimisation.

There is no formal way to begin an evolutionary optimisation process, however Bentley (1999, p. 36) state that it will typically involve an existing design artefact to be first represented computationally. In such process the designer is to decide on which aspects of the design artefact need to be improved and express those in variables. This is the process of parameterising design artefact parts or assembly instructions into variable genes is called genetic coding. (Bentley, 1999, p. 429) However, this should not be mistaken for term Parametricism used by Patrick Schumacher (2009) to describe the on-going use of parametric design as a unique style, but this will addressed in the last chapter of the paper. Often genotype to phenotype mapping is not required as the parameters could be directly applied to the solution artefact. (Bentley, 1999, p. 37) At this stage it is important for a designer to specify what the selection criteria for the solutions is and how fitness values of the phenotypes will be calculated. To explain this type of evolutionary optimisation Bentley (1999, p. 37) uses a four-legged table (fig 4.). He enumerates the distance from the centre of the table to its four corners for each table leg and their lengths. In this case the phenotype consists of 8 values, length and distance from the centre for each leg, and genotype being 64 bits. After a random initial population is evaluated, the increase of the distance parameters and decrease of legs’ lengths will eventually improve the stability of the table, however the designer could imply limitations by specifying the minimum and maximum values allowed.

Figure 4: Evolutionary optimisation of a table

It should be noted that even though the resulted solutions are to single design artefact as a table the EA will evolve a population of phenotypes to satisfy the selection criteria. Therefore, the result will consist of variation of unique possible solutions similar to each other ,just as in nature, but not identical. Such structural optimisation problem can easily be solved by EA as the genetic values are not necessarily in conflict with each other, thus a population of nearly optimal solutions to a single criterion can easily be achieved in small number of generations.

Nevertheless, an evolutionary optimisation solution could often need to satisfy multiple criteria sometimes even conflicting with each other, as shown by Alexander. In such case the multiple conflicting criteria will balance each other thus creating more inhabitable environment for evolving populations. It could be said, that the search of the solution space in cases of evolutionary optimisation is being done under different terms or rules. In example of single selection criteria, the rules for the evolutionary search are strict, meaning there will be limited number of parameters to change in order to achieve a fittest result for a specific selection criterion. Whereas in the case of combined multiple selection criteria, the rules of search are becoming rather loose, meaning that the evolved population can consist of greater variety of phenotypes due to the evaluation environment being more inhabitable. As a result of the more inhabitable environment or balanced selection criteria, extreme transformations can occur. The designer could end up with number of solutions meeting the requirements, but all of them adapting to different parts or niches of the environment. This could be seen as a process of speciation or diversity. To better demonstrate this, an analogy could be made with a complex system such as a tropical rainforest. Holland (2014) states that a tropical rainforest can host almost endless variety insect species, ranging from extreme generalist, like army ants able to feed from almost anything, to extreme specialists like Darwin’s ‘comet orchid’, which could be pollinated by specific species of moth due its foot-long nectar tube. In the analogy, the army ant would be the ideal solution to problems of optimisation, thus being best performing solution to number of requirements. Whereas, the comet orchid being likely to be found in less common areas of a solution space would still be a possible solution, however it will exceed in performance for certain requirement but perform poorly in others. According to Steadman (2008, p. 255), in some evolutionary design systems, multiple fitness criteria are combined according to the principle of Pareto optimality, borrowed from classical economics. In the present context, a solution is Pareto optimal if performance cannot be improved on any one criterion without worsening performance on others, as the described in the case of the army ants. In case when multiple solutions mathematically meet a define selection criteria any favouring one solution before another would be highly subjective and based on designers’ intuition and preferences.

A practical example of such inhabitable environment can be seen in the Design Explorer tool created by CORE Studio (Howes, 2016) (fig.5) where the selection criteria are enumerated separately and acting as filters. As mentioned earlier, EA evolve and populations rather than single individuals so in case where number of individuals (army ants and Darwin’s comet orchid) a single optimal solution could not be given as they both meet the global selection criteria. In the given case, the Design Explorer allows the all separate selection criteria, such as depth, height, orientation, heating, cooling and other, to be manually changed by a designer thus making the exploration of the solutions space personal and ill defined.

Coming in to this chapter the paper took the architectural design problem of optimisation or finding fittest solution to set of requirements as described by Christopher Alexander (1964, p. 15) in his Notes on the Synthesis of Form. It then presented examples how a design artefact could be optimised by EA being used to evolve population of improved solution. By enumerating the parts of the design artefact, a designer allows the EA to evolve solutions by testing them to predefined selection criteria, thus automating and speeding up typical trial and error process. This confirms the second hypothetical point that evolutionary processes can be used in architectural design with optimisation orientation.

Figure 5: Design explorer by Core Studio

Although the above may be true, the process has its limitations. First, Steadman (2008, p. 251) warns the designers that EA should not necessarily be used if an established analytical technique to directly determine a design solution is already available. Second, in the case of combination of multiple selection criteria EA can still evolve improved population, but it cannot specify single best performing or optimal solutions, for this reason the decision is best left to the designer. However, as a result of EA being used to evolve populations, the fitness values of individuals should will also be available to the designer. Thus, any critical decisions a designer makes could also be supported with unbiased performance information of individual solutions. Before concluding the question of optimisation, it should be said that, EAs have a kind of natural generative properties as they evolve number of individual solutions in a population. This property of EA work with populations and create diversity can contribute to the last point raised to be investigated in the paper – use of evolutionary processes with generative orientation. However, as seen in the Bentley’s example, the optimised table design did not radically change its form and style. Therefore, this kind of given generation will be questioned further in the parts to follow.

Problem of Generation

Considering that one agrees with the described use of evolutionary algorithms for design optimisation, it is now turn for last point of the hypothesis to be addressed – evolutionary design as a generative tool.  It was mentioned briefly when discussing CAD tools that Frazer (1995, p. 60) defines the present architectural design processes as fundamentally wrong and not worth imitating. When discussing his proposal for evolutionary architecture he then writes that: “A clear distinction is intended between sources of inspiration and sources of explanation… In this context, ours [as architects and designers] in not the theory of explanation, but a theory of generation.” (Frazer, 1995, p. 12) It could be argued that many architects similarly define generation as a design problem in architecture. Thus, the investigation shall continue to by looking at the emergent properties of EAs to provide novel architectural design artefacts. Since Frazer raised the enquiry is it reasonable to begin with his theoretical propositions

In a way Frazer (1995) not only takes Alexander’s idea of an ensemble but adds on to it by defining evolutionary architecture as a metabolism or an adaptive system. Thus, he allows the design artefact to transform by responding to contextual environment it is situated in. Frazer refers to the architectural design as a process of grown or development of biological organism. Similarly, one of the four aspects of evolutionary design by Bentley was evolutionary artificial life forms. His proposal for achieving evolutionary architecture requires embedding the physical properties of a design artefact into variable parameters and allowing it to grown thought iterative process by interacting with unstable/continuingly changing environment. In his suggested design process the architect provides an explicit “blueprint” or set of instructions for the potential development of the design artefact. (Frazer, 1995, p. 11) The architect is aware of his intentions and the goal he wishes to pursue but he is ‘blind’ to the outcome of the development process. In traditional architectural design the architect has full control over the production of the design artefact. Here, on the contrary, the architect is only in control over the initial concept, or as Frazer used the term seed coming from botany, but not in its later development of the design artefact. Speaking biologically the architect defines the genotype, the core properties of the design artefact, but its phenotype development depends of the contextual environment it is exposed to. It should be added that the biologically inspired computational methods Frazer describes are not only coming from evolutionary side of biology, but also from the development side.  Thus, there is important distinction to be highlighted between the evolution and development. Evolutionary processes investigated in this paper as described in the beginning are coming from the modern synthesis of evolutionary biology resulted from the combination of Darwinian ideas with Mendelian inheritance. The development deals with how organism interact with the habitual environment and how their growth is affected by this interaction. The way Frazer uses evolutionary processes is to continuously evaluate the fitness of his design artefacts to their environment. In evolution, nature act as a hostile environment, thus populations generally improve one generation after another by investing in well-performing individuals and allowing worse-performing ones to die. This property of organisms to continuously improve and adapt to hostile environment, described by Holland (1992), is how Frazer uses GA to stimulate a feedback response in his metabolic architectural artefacts to changing environment.  In simple terms, in order achieve growth and development similar to living organisms in his architectural design artefacts, Frazer combines developmental systems with GA.

An example of this combination could be seen in Frazer’s Universal Interactor at the Architectural Association in London. (Frazer, 1995, p. 75) It is said for the Universal Interactor that it takes input sound, wind and sun signals from the environment via antennae. The signals were then used in a simple environmental model formulating response based on an EA. The developmental section, in its turn then processes responds to the signals (Frazer, 1995, p. 78) Examples of the forms produced with this interaction can be seen below (fig.6).

Similar to previous examples, a designer has control over the instructions and rules to be used in the evolutionary design process, however this time a computer is given extra freedom to suggest design solution. In this way evolutionary design is used as a tool assist in the generation of possible solutions, allowing the designer to speculate over their potential implementation in architectural practice. Therefore, one could reasonably see the combination of EA with developmental models to produce variety of design artefacts as another way to use evolutionary processes with generative orientation.

Figure 6: Digital models grown using the Universal Interactor

Similar to Frazer, Ahlquist and Menges (2011, p. 21) emphasize of the emergent properties of evolutionary algorithms – their ability to evolve solutions without strictly defined problems. They see a design problem or more precisely the framework of a given brief as an opportunity to step away from conventional design methods, to speculate and innovate. The important aspect on evolution for Ahlquist and Menges (2011, p. 21) is not the embedded optimisation power on the natural processes, but the open-ended characteristics. By continuously improving the short-term survival of organisms, within the three Darwinian rules, nature constantly pushes the boundaries of the real-world solution space and challenges the Newtonian paradigm of stability. (Frazer, 1995, p. 13)

A Pneumatic Strawberry Bar (fig.7) designed by Menges (2003) aims to extend use of evolutionary processes of reproduction, mutation, competition and selection as design strategies. For this reason, the Strawberry Bar seems as a good candidate to give further examples of how evolutionary processes are used with generative orientation. The project is closely bound to the manufacturing process, in this way it bridges the gap between form-generation and fabrication. The physical properties of the fabrication material were incorporated into the evolutionary design process most likely in a fitness function, thus being used calculation of the fitness values of individual solutions. Menges (2007) describes the project as consisting two trapeziform surfaces aligned at a fixed starting point. The inflated component is a three-dimensional form defined by the different length of the surfaces in relation to the defining points and the spatial origin. These simple geometric relations, defined as a generic 3D cutting-pattern, provided the base data for the subsequent process that simultaneously grew three sub-populations of surfaces. After running the evolutionary process over 600 generations Menges (2007) reported that 144 species were identified and catalogued according to specific patterns of relevant geometric features. Due to the interrelated evolution of the geometry-defining surfaces the criteria for evaluation was the relative fitness amongst the emergent species rather than the absolute fitness ranking of any particular individuals. Without discussing the success of the project, it should be elaborated how Menges arrived at the final result.

Figure 7: Strawberry Bar, Morphogenetic Design Experiment, 2003

For the design of the Strawberry bar Menges used the tool called Genr8. (Menges, et al., 2007) For Genr8 is said that to allows both growing and evolving three dimensional digital forms or surfaces by an architect or designer as the tool combines L-system with EA. (Menges, et al., 2007) Lindenmayer grammar (also known as L-grammar or L-system) is a rule-based system for representing the topology of the branching structures of plants by means of symbols. A string of symbols represents the initial state, and the rules specify how symbols are to be replaced recursively with symbol strings. The symbols can in turn be made to code for parameters such as the lengths and angles of branches and the shapes of leaves and flowers. (Steadman, 2008, p. 257) Generally speaking, in biological context the L-systems comes from developmental processes of organisms rather the evolutionary branch. In this case, both the Universal Interactor by Frazer and the Strawberry Bar by Menges are not using evolutionary processes independently as a generative tool, but in combination with developmental processes. Perhaps, due to the Strawberry Bar being directly evaluated for its structural performance use of EAs is not surprizing, but the combination evolutionary and developmental processes instead of their individual use certainly is.

There are more examples of such combination of evolutionary and developmental processes being use in architectural design. The list the includes the Universal Constructor at the Architectural Association in London by John Frazer (1995, pp. 44-45) using cellular automata and the exploration of three-dimensional worlds by Paul Coates (1999, p. 323) using  Lindenmayer system and Genetic Programming. On the question why using combination of L-Systems and Genetic Programming, Coates (1999, p. 323) states that when EAs are used for optimisation the fundamental design decisions regarding any design artefact were already taken. After all, the selection criteria in EAs regardless of being strict or loose, as described previously, remains constant throughout the evolutionary search. Under those circumstances, one could argue that Frazer, Menges and Coates saw this as a limitation of EAs and decided to combine them with developmental systems in order to achieve higher unpredictability and novel solutions.

After confirming the optimisation properties of evolutionary algorithms the paper explored their use as a generative tool leading to novelty in architectural design.  Two examples were presented – the Universal Interactor by Frazer and the Pneumatic Strawberry Bar by Menges to demonstrate how EA are used in the production of novel and radical design artefacts. Regarding the last hypothetical point, following the given examples one can conclude that evolutionary processes certainly can and in fact are used with generative orientation. Although, it seems like their emergent properties may not be as strong as their optimisation potential, thus some of the experts in the field have combined them with developmental process.

Commentary on the findings

On the question of computers in design, Alexander dismissed their use stating that there very few design problems that can be given to a computer. (1967, pp. 8-12) Instead of looking at already established problems in architectural design the more appropriate question should have been how to define a design problem, so it could be given to a computer. After all, the computers are one of the great discoveries in recent history and as such they should be used more efficiently in architectural design. Due to the close analogies being made between biology and architecture it has was speculated that evolutionary processes can allow computers to automate design tasks as well as being used for optimisation of design artefact and generation of novel solutions. To address the first point, the paper outlined the mechanics of evolution coming from the modern synthesis of evolutionary biology as described by Dawkins. It then presented that evolutionary processes can be encoded in evolutionary algorithms and given to a computer to process. Presenting that computers can be given evolutionary tasks was perhaps the most important finding of the paper. If the given methods of using computers in architecture becomes more popular this could result lead to a decline of traditional drafting and modelling methods. This is not to say that computational design is superior to hand crafting. Use of computers as tool for automation can certainly open more doors for design exploration, but it should not replace traditional design methods entirely. Thus it is individual architects’ responsibility to establish a balance between both. As a contribution to the on-going discussions of artificial intelligence in design, the paper presented that by using evolutionary methods designers still have great amount of control over the machines by having an input at every stage of the evolutionary design process. It should be noted that the paper was not intended to be overly technical, thus in-depth description on the biological and computational aspects was left out can certainly be found if needed.

Going to the second point raised at the beginning, the paper looked at optimisation of design artefacts. The optimisation problem was reviewed partly because of Alexander (1964, p. 2) statement that it is typical problem in design and his rejection of computers and also due to personal experience in practice. Evolution as described by Darwin is process of gradual change (improvement), therefore by evolutionary process successfully recreated virtually it should be no surprize they can be used for optimisation. Perhaps the more important fact here is that in cases when multiple conflicting criteria are to be satisfied an EA cannot give a single best solution. As per the given the example, in such cases a population of optimised solution is generated, but it is the designer responsibility to make the critical decision of choosing a specific solution. Some architects would prefer being able to make the final decision, whereas others wishing to be given a single solution would perhaps find this as a limitation of the EA. Nevertheless, the general ability of EA to optimise design artefacts is certainly evident. A designer is given the opportunity explore the solution space by changing to selection criteria after the initial requirements have been met. The paper described to type of evolutionary searches – strict and loose. The two terms refer to evaluation environments (criteria) being hostile or inhabitable, but when used to describe the evolutionary searches the proposed terms seem rather appropriate.

For many architects the search of novelty if of great importance. For this reason, the last point raised by the paper was the use of evolutionary processes with generative orientation or as a tool for design generation. Since Frazer presented this issues in his book, it was appropriate to first look at his work and propositions. The paper did not specify the term novel as it was intended to be used in a broad definition to promote creating exploration of a solution space. As seen from earlier examples, EA deal with populations – group of improved solutions. Seen in this way EA already have an evident emergent property resulting a variety of improved solutions. However, it seems that some architects find this as not sufficient and preferred to combine evolutionary processes with developmental ones. On the topic of emergence Holland (2011, p. 141) made an interesting observation saying that rules that are absurdly simple can generate coherent, emergent phenomena. Perhaps this could be the key when using EA with for generative purposes, the criteria should be as loose as possible to allow for greater variation to occur. Seen evolutionary processes being combined with developmental ones in the work of Frazer, Menges and Coates tease the potential of such developmental processes and this could be a future research as a continuation of the current paper.

All things considered, evolutionary processes seem to be every useful when employed for optimisation purposes and promising when used for generative ones. With EA architects can automate trial and error design process previously described as being to lengthy and allowing only sound options to be extensively explored. Given this effective use of EA, one could reasonably conclude that evolutionary design is rather appropriate to be used in practice and should be given greater attention. Before finalising the paper some of the more common limitation and misconceptions about evolutionary design should also be acknowledged.

Section IV

Common limitations and misconceptions of evolutionary design

The paper so far exhibited the use of evolutionary processes in architectural design with both optimisation and generative orientation. In general, EAs seem robust enough to be used in architectural design an alternative to traditional of-the-shelf CAD software. However, one could reasonably question why such propositions have not already been implemented in practice. The following parts of the paper shall present some of the assertation holding back evolutionary design tools from being widely accepted.

The first question many would probably ask is why evolutionary algorithms are not used more in practice if they are so universal and robust? Computers, just like EAs, are relatively quite recent invention in historic terms, but their impact on contemporary living is undisputable. For this reason, the paper shall take a step off topic to make an analogy with cellular automata which was studied during the course of production of this paper. In his book A New Kind of Science Stephen Wolfram demonstrates great variety of CA patterns generated through iterative process driven only by discrete set of very simple rules. Wolfram (2002, p. 42) argues that these experiments have not been undertaken in the past. The technological limitation and time involved, if done by hand, are certainly an important factor, the most likely reason for that nobody made this discoveries before seems to be the fact that people oversaw the emergent properties of CA and the intricacy of the generated patterns. Wolfram (2002, p. 42) reasonably argues that the intricate results of his experiments could easily be left unnoticed, as standard intuition gives no reason to think otherwise. As mentioned before, Darwin’s theory of evolution consists of only 3 simple rules. Despite the simplicity of the rules of biological evolution, they were left undiscovered by thinkers of the calibre of Galileo, Newton, Aristotle.  In fact, nobody before the mid nineteenth century made such observation. (Dawkins, 2006) Even after his theory was published it faced great criticism, much of which were arguments from personal incredulity. Perhaps for similar reasons evolutionary design has not yet became widely used in practice. First, traditional CAD tools are predominant in the design and manufacturing, leaving little space for other alternative tools to be introduced. And second, as seen the origin of evolutionary design might seem by one as too specific for such a universal tool. Of course, if evolutionary design does become part of the architectural curriculum, then it would be reasonable to expect more examples of its practical use in future.

Another fall back might be the technical side of evolutionary design and the resulted increase of information. As seen previously does require some knowledge of computation and mathematics in order to be programmed. Typically, many architects do not have great exposure to the field of computer science, or if they do, some decide to stay away from this new medium. In addition to computation, architects would need to switch often between two very different tasks: creative design exploration and formal mathematical representation of the design artefact and its contextual environment. According to Aish and Woodbury (2005), the parameterization involved in the parametric design increases complexity as the designer is required to work on both the artefact representation and the framework within which variation can occur. In other words, evolutionary design does not reduce complexity in contemporary architectural design. Instead, it is perhaps the opposite, imposing more knowledge to be assimilated by designers. Steadman (2008, p. 251) also warns that EAs should not necessarily be used if an established analytical technique is already available.  It’s then left to a designer to distinguish tasks worth being externalised to a computer and tackled in the digital realm from the ones that are not.

As seen in the previous parts of the paper, working with evolutionary algorithms requires some design characteristics to be represented in a genetic form, more specifically in variables. However, the term parameterisation refers to simply enumerating or encoding the characteristics into variables in order for a computer to be able to processes the given data. Ahlquist and Menges (2011, p. 19) make a step further from the conventional understanding of parametrics as interdependencies of geometric constrains. They argue the idea of parametrics goes to establishing mathematical rules as representation of behaviours and forms. In the past 20 or more years parametric tools such as Grasshopper for McNeil Rhinoceros, Dynamo for Revit and others became accessible. The design artefacts produced with these tools often share similar aesthetics. Parametric architecture is often characterised with free-form and organic design, however it is rather ignorant to define parametric design as a unique style as described by Patrick Schumacher (2009) and his Parametricism. This leads directly lead us yet another misconception –  evolutionary design does not equals complex geometry. As presented in the paper, it is necessary for variables and parameters to be used in evolutionary design, but the resulted forms could be made rigid and inorganic (contrary to the term organic architecture). The described fluid aesthetic outcome is rather consciously pursued by individual designers and certainly not a result of the specific use of evolutionary or other computational tools.

Conclusion

The paper began by asking if computers have the capacity to perform design tasks without being explicitly programmed how to do them, but being given what is to be achieved? It then looked at biology and specifically evolutionary processes and presented examples how these could be recreated virtually and be used in architectural design with optimisation and generative orientation. The examples show how the designer could deploy parts of its intelligence to the machine and give it a level of autonomy to process the required data. Given promising results and robust qualities of evolutionary algorithms one could reasonably conclude that computers should be addressed more often in this manner, thus stepping away from traditional use of computers as digital drawing boards. Then the investigation went on presenting some of the common limitation and misconception of evolutionary design by computers and suggested why it may have not been used by practicing architects. Regardless if one decides to employ evolutionary processes in architectural design they should not be ignored.

We must face the fact that we are on the brink of times when man may be able to magnify his intellectual and inventive capability, just as in the nineteenth century he used machines to magnify his physical capacity. Again, as then, our innocence is lost. And again, of course, the innocence, once lost, cannot be regained. The loss demands attention, not denial. (Alexander, 1964, p. 11)

Bibliography

  1. Ahlquist, S. & Menges, A. (2011) Computational Design Thinking. In: S. Ahlquist & A. Menges, eds. Computational Design Thinking. l.:Wiley.
  2. Aish, R. & Woodbury, R. (2005). Multi-Level Interaction in Parametric Design. Berlin, Springer.
  3. Alexander, C. (1964) Notes on the synthesis of form. Cambridge, Massachusetts and London, England: Harvard University Press.
  4. Alexander, C. (1967) The Question of Computers in Design. Landscape, Issue Autumn, pp. 8-12.
  5. Bentley, P. J. (1999) An introduction to evolutionary design by computers. In: P. J. Bentley, ed. Evolutionary design by computers. San Francisco, California: Morgan Kaufmann Publishers.
  6. Coates, P. (2010) Architecture. Abingdon: Routledge.
  7. Coates, P., Broughton, T. & Jackson, H. (1999) Exploring Three-Dimensional Design Worlds using Lindenmayer Systems and Genetic Programming. In: P. J. Bentley, ed. Evolutionary Design by Computers. San Francisco, California: Morgan Kaufmann Publishers, pp. 323-341.
  8. Dawkins, R., (2006) The Blind Watchmaker. London: Penguin books.
  9. Frazer, J., (1995) An evolutionary architecture. London: Architectural Association.
  10. Gero, J. & Rosenmann, M., (1999) Evolving Designs by Generating Useful Complex Gene Structures. In: P. J. Bentley, ed. Evolutionary design by computers. San Fracisco, California: Morgan Kaufmann Publishers, pp. 345-364.
  11. Gero, J. S. (1994) Preface. In: J. S. Gero & E. Tyugu, eds. Formal Design Methods for CAD. Amsterdam: Elsevier.
  12. Holland, J. (2011) Constrained Generating Procedures. In: S. Ahlquist & A. Menges, eds. Computational Design Thinking. l.:Wiley, pp. 131-141.
  13. Holland, J. H. (1992) Adaptation in natural and artificial systems. 1st MIT Press ed. Cambridge, Massachusetts and London, England: MIT Press.
  14. Holland, J. H. (2014) Complexity: A Very Short Introduction. Oxford: University Press.
  15. Howes, B., (2016.) Design Explorer – Announcement. [Online]
    Available at: http://core.thorntontomasetti.com/design-explorer-announcement/
    [Accessed 01 April 2018].
  16. Lamarck, J. B. (1916) Histoire naturelle des Animaux Sans vertèbres. Translated from French by Russell, E.S, In: Form and function : a contribution to the history of animal morphology. London: J. Murray, p. 221.
  17. Menges, A. (2003) MDE 02 AA Strawberry Bar. [Online]
    Available at: http://www.achimmenges.net/?p=4434 [Accessed 22 April 2018].
  18. Menges, A. et al. (2007) Genr8: Architects’ Experience with an Emergent Design Tool. In: J. J. Romero & P. Machado, eds. The Art of Artificial Evolution: A Handbook on Evolutionary Art and Music. l.:Springer Berlin Heidelberg, pp. 167-188.
  19. Parmee, I. (1999) Exploring the Design Potential of Evolutionary Search, Exploration and Optimisation. In: P. J. Bentley, ed. Evolutionary Design by Computers. San Fransicso, California: Morgan Kaufmann Publishers, pp. 119-143.
  20. Schumacher, P. (2009) Parametricism – A New Global Style for Architecture and Urban Design. AD Architectural Design – Digital Cities, July/August, 79(4), pp. 14-23.
  21. Steadman, P., (2008) The evolution of designs: biological analogy in architecture and applied arts. Revised ed. Abingdon: Routledge.
  22. Terzidis, K., (2011) Algorithmic form. In: S. Ahlquist & A. Menges, eds. Computational Design Thinking. l.:Wiley, pp. 94-101.
  23. West, G., (2017) Scale: The Universal Laws of Life and Death in Organisms, Cities and Companies. London: Weidenfeld Nicolson.
  24. Wolfram, S., (2002) A New Kind Of Science. First ed. s.l.:Wolfram Media Inc.