Digital code
scripts for generative and evolutionary design: De Identitate
Professor John Hamilton Frazer, AA
Dipl, MA, FCSD, FRSA
Autotectonica
With additional contribution by Patrick
Janssen
Abstract
The question of De
Identitate or identity is addressed by describing the notion of concept-seeding
which encodes design characteristics in DNA like code script with generative
and evolutionary capability.
Introduction
Generative design
systems are used to
generate large numbers of design alternatives with significant differences between them. These systems define a complex growth process to transform an encoded seed into a design. By making
small modifications to either the transformation process or the seed,
alternative designs can be generated.
Evolutionary design
systems are used to evolve
designs adapting to their
environment, loosely based on the
neo-Darwinian model of evolution through natural selection. These systems consist of a cyclical process of continuous manipulation to ensure the population of designs evolve and adapt
gradually.
(Frazer and Janssen 2003)
Three techniques are described in this paper:
Concept-seeding
proposes a system generating
environment-responding designs. This invokes a set of architectural concepts captured and codified by the designer. Through small
command modification, alternative designs are generated.
The
generative-evolutionary model makes use of an evolutionary design system
embedded within a generative system. The generative system generates
alternative designs in response to environment, and the
evolutionary system manipulates and evolves the generative modifications. In this case,
the system does not use any codified architectural concepts.
The combined
model, combines the generative-evolutionary system and the concept-seeding
approach. It allows the generative-evolutionary system to evolve alternative designs, embodying the
codified architectural concepts.
De
Identitate
Most designers
employ a highly personalised methodology. This is often apparent stylistically such as with Gaudí, Mackintosh or Frank Lloyd Wright. Or it can be
seen more organisationally or procedurally, or be concerned with abstract space
and form. This gives the designer their identity. And this can also
apply to whole cities or even cultures up to a point.
This personalised but generic methodology is very abstract, but it can be manifest as sets of standard details formalised to ensure a
consistency and to reinforce the style, which can be easily
evidenced inside architect’s office. The
identifying characteristics go through changes during the development of the
designers, sometimes with abrupt changes, more usually with a continuous progression; and even more
interestingly, it can continue long after the death of the original
designer
(Frazer & Janssen 2003)
If the designers’ architectural concepts could be captured and codified
in a generic form, a generative system might be able to invoke them to generate
designs to embody the concepts. This approach of capturing and
codifying architectural concepts is referred to as concept seeding (Frazer, 1974, 1979).
Three tasks are
defined. First, a set of generation rules are defined to develop the concept seed into a design. Second, a
concept seed is defined to capture certain architectural concepts. Third, with a generative system, designs are generated in response to the environment, both the context and the criteria. The tasks of model’s identifying are not mutually
independent, in parallel in most cases.
The architectural
concepts can also be codified historically as the endless
attempts to recapture the nature of the paintings of Mondrian or the villas of
Palladio. The success or failure varies and depends
on the sensitivity of the re-creator, often with crass results, particularly when trying to create new works by dead artists. In the case of some living designer’s involvement, such as our work with Cedric Price and Walter Segal, whose timber housing system has been captured for perpetuity in the software (Frazer, 1982).
The concept seeding system is, in itself,
not a cyclical system. The system generates a single design proposal from a
single seed responding to
the design environment. The
designer will
explore a wide range of design possibilities with small generative modifications to either
the concept seed or the generative rules, therefore, resulting in a cyclical process guided by the designers.
The Reptile System
The first attempts
to realise this approach was the Reptile
System, developed by Frazer from
1968 onwards (Frazer, 1974). This system was
developed as a folded plate space-frame system, capable of creating a wide variety of enclosures from
two basic structural units, which can be orientated
in 18 different ways relative to each other, resulting in over three hundred useful combinatorial
possibilities.
As drawing
a Reptile enclosure by hand was very tedious, a computer program was developed to draw the enclosures and to create perspective views, initially in 1967. At this stage, with extremely limited capabilities of the computer hardware, much effort focused on how to store and manipulate these enclosures. As the program was enhanced with additional features, by 1971, a generative program was developed to
semi-automatically generate complete Reptile enclosures.
Generalised
and extended version
A generalised version of the program based on component was later
developed. Two kinds of information were required in the program, the conceptual model of the building
information in its minimal coded configuration, and a description of the actual components and
details for the output stage.
The concept of
cultivating the seed to produce different buildings has been extended to mutate details of the seed, producing variants to
overcome the great increase in environmental variety encountered with more
general purpose of building
construction. Individual mutations of the assemblies can be developed
interactively and stored as variants.
The configuration of the
minimal construction or seed is compared with the brief for the building, and the seed is grown, stretched, deformed and pruned,
until it conforms to these requirements. The same seed may bring the variety of forms of building to the same
requirements, same as form of building being produced by different cultivation routines. The data structure is
modified in two ways. First is the optimisation of the quantifiable and specific brief. Second is the evaluation of solutions produced
by the first technique, including aesthetic judgements (Frazer, 1979).
Generative
evolutionary model
The symbiotic behaviour and
metabolic balance of the natural environment are attempted by the generative-evolutionary approach in the built environment. This model proposes the natural evolutionary process to generate the architectural form. The prototype and creative power of natural evolution are emulated
by generating virtual architectural designs responding to changing environments. Like the natural world, architecture as an artificial life, subjects to principles of morphogenesis, genetic coding,
replication and selection (Frazer, 1995).
A generative system uses
code-scripts of instructions to produce computer models of alternative designs
to simulate the development of prototypical forms, then, evaluated on the basis of
their performance in a simulated environment. By mutating and manipulating the
code scripts, new forms are generated, sometimes with unexpected forms emerging.
Typical
evolutionary approach
Most engineering
applications of the evolutionary approach are interested in convergence on an
optimal solution to defined computable criteria for selection. This model
identifies two tasks: codifying the parametric model and evolving designs in
response to the environment. There are two steps in the first task, one is mapping rules, to map the parameters as a model, and second is evaluating rules to evaluate the model.
A mapping rule
produces the design from an encoded set of parameter values by inserting the
values into the parametric model. The evolutionary system evolves these
parameter values. We define this as convergent
evolution by natural selection.
As all the designs are based on the same parametric model, all of them will have the same overall organisation and
configuration. Therefore, these programs
are seriously limited in their ability to evolve new designs.
However convergent
evolution by natural selection is not the only possibility. In the Origin of
Species, Darwin talks first of the technique of artificial selection. In
our model artificial selection opens
up the opportunity to demonstrate preference for user concerns, and provides the
opportunity for the designer to jump to faster results.
Nature also relies on divergence to keep a varied gene pool for sudden changes in the environment. Our model provides divergent evolution for the generation of alternative ideas. This gives a matrix of four possible combinations of natural/artificial selection and convergent/divergent evolution (Dawkins, 1986; Simms, 1991; Graham, 1993; Frazer, 1995). However there is no need to slavishly follow nature and short cuts can be made in artificial evolutionary systems such as allowing Lamarckian inheritance.
Generative-evolutionary
approach
A
generative-evolutionary system replaces the mapping step with a generative
step. This generative step consists of an embedded generative system. The generative-evolutionary approach requires describing the generative rules in a genetic code in response to a simulated environment. The generative rules tend to be general, without intention to reflect particular
architectural concepts. All such rules contain biases
and constraints, and the forms produced
may nevertheless all share certain characteristics.
The model
identifies two tasks: codifying the generative concepts and evolving designs. First, the generation rules
generate designs from encoded code-scripts and the evaluation rules evaluate
the generated designs. Second, designs are evolved in response to the design
environment by a
generative-evolutionary system.
Initially, a
population of code scripts is created. The evolutionary process consists of
four steps. First, designs are generated from the genetic code scripts through
some form of epigenetic development in an environment. Second, designs are
evaluated by simulating and analysing their performance within their
environment. Third, the most successful are selected. Fourth, the selected code
scripts are transformed by genetic operators such as crossover and mutation.
These four steps then repeat. At some point the process may be stopped.
An important aspect
of the generative system is to generate designs in
response to an environment. The environment consists of the informative design context and
criteria to
the generative rules. This allows the developmental process of design as an adaptive process, determining successive structural
modifications in response to the environment and measuring the performance of different structures in the
environment.
The Interactivator
In 1995, a
generative-evolutionary experiment was launched to involve global participation
in the evolution of a virtual environment, the evolutionary
model of nature was proposed as the
generating process for architectural form.
The model proposed a generative-evolutionary system accessible via
the Internet, to encourage wide participation and to create biodiversity
in the genetic design pool on which Janssen’s model was involved in developing a special
demonstration version, the
Interactivator. With simplification of the theoretical system, all the key
elements are represented.
The developmental
process of each member of the family consists of three cyclical parts: cellular growth, materialization and the genetic
search landscape. A genetic algorithm is used to ensure that future generations
learn from the previous ones and provide for
biodiversity during the evolutionary process.
Combined model
The concept seeding model combined with the
generative-evolutionary model results in a new type of
model with both advantages of previous models. The concept-seeding model
allows generating
designs of particular architectural concepts. The
generative-evolutionary model allows designs to evolve, adapting to their
environment in complex ways. The combined model synthesises these two previous
models, thereby allowing designs to evolve in response to their environment and to embody particular architectural concepts.
In the case of the
concept seeding model, the combined model evolves the generative modifications
manually by
the designer. These modifications
introduced small changes to either the seed or to the rules. Representing these
modifications allows an evolutionary system to evolve the modifications. The
generative modifications are used to generate designs to be evaluated, based on which new populations of modifications are created. This
approach requires defining the range of valid
generative modifications to develop rules, automatically creating new generative modifications.
With the previous model,
generation and evaluation rules are defined. The concept seed is defined, codifying a set of
architectural concepts. The design alternatives are
evolved in response to the design environment with a generative-evolutionary system and concept seeding.
The embedded
generative system generates designs from concept seeds. The generated designs
will embody the set of architectural concepts codified by the seed. These
modifications will make small changes either to the seed itself or to the
rules. These generative modifications result in different designs, with the highest fitness scores being selected. A new population of
generative modifications is operated to generate a new population of designs, and so forth.
Janssen’s
combined model
Broadly speaking,
Janssen’s model follows the pattern developed in Frazer’s combined model, which focus on two areas: the nature of the codified concepts
and the structure of the overall model.
The idea of capturing and
codifying the architectural concept has been refined in a number of ways, which is aimed to develop creative architectural concept. Janssen has defined ‘variety’ more precisely by using a distinction made in evolutionary biology
between diversity and disparity (Jaanusson, 1981; Runnegar, 1987; Gould, 2000).
Diversity refers designs different in the proportions
and dimensions of their parts, with the same overall
organisation and configuration of parts in common.
Disparity refers to designs that have a fundamentally different organisation
and configuration of parts.
The architectural
concept must contain enough flexibility and adaptability to allow the birth of disparate designs. The architectural
concept should not predefine the overall organisation and configuration of the
designs, instead, should focus on defining the
parts of the design and their interactions and overlaps. These parts,
interactions, and overlaps might be thought of as defining the character of the
designs without specifying their overall form. Janssen therefore refers to the
set of architectural concepts to as the character schema (or in the
context of this conference: De Identitate)
Splitting the
design procedure into two phases emphasises the fact that these two phases can
be executed in very different ways. The first phase develops, and codifies the
design character. This character will reflect the beliefs and preferences of
the design team – called the design stance – and will be developed in
response to the niche environment. This niche environment can be defined before
any specific design environment has actually been found. As a result, this
phase creates a generic design entity – the evolution schema – that can be
reused many times within different projects.
Conclusion
The approach
implies some changes in architects’ working methods. The generic approach has
to be made explicit, rigorous, and stated in terms which enable a concept to be
expressed in genetic code. Ideally, the computer could deduce this information
from normal work methods without any conscious changes being necessary.
Architects have to be clear about the criteria for evaluating an idea, and
prepared to accept the concept of client- and user-participation in the
process. The design responsibility changes to one of overall concept and
embedded detail, but not individual manifestation. Overall the roles of the
architect is enhanced rather than diminished as it becomes possible to seed far
more generations of new designs than could be individually supervised, and to
achieve a level of sophistication and complexity far beyond the economics of
normal office practice.
Interactivator: Networked Evolutionary Design System
John Frazer, Julia Frazer, Manit
Rastogi, Peter Graham, Patrick Janssen
Architectural Association, London, 1995
Acknowledgements
This paper is based on work by
the author and his students at the Architectural Association in London,
Cambridge University, the University of Ulster and the School of Design in Hong
Kong. The valuable contribution of Patrick Janssen is particularly acknowledged
– he worked on the Interactivator in 1995 and provided the Janssen
model described at the end of the paper.
The author would also like
express his gratitude to GU Yan for substantial assistance in preparing and
editing this paper based on lectures given by the author at the Bartlett School
of Architecture, University College, London, 2004, and on a publication by
Frazer & Janssen (2003) and other papers.
References
Dawkins, R. (1986), The
Blind Watchmaker, Longman.
Frazer, J. H. (1974),
Reptiles. Architectural Design, April, pp. 231 - 239.
Frazer, J. H. and Connor, J. M. (1979), A conceptual
Seeding Technique for Architectural Design. PArK 79, Proceedings of the
International Conference on the Application of Computers in Architectural Design.
(Online Conferences with AMK, 1979), Berlin, pp. 425 – 434.
Frazer, J. H. (1982), Use
of Three-Dimensional Computer Input Devices to Encourage Public Participation
in Design. Proceedings of the conference on Computer Aided Design 82, Butterworth
Scientific, pp. 143 – 151.
Frazer, J. H. (1990), A
Genetic Approach to Design – Towards an Intelligent Teacup. In The Many
Faces of Design, Nottingham.
Frazer, J. H. (1992),
Datastructures for Rule-Based and Genetic Design. In Visual Computing –
Integrating Computer Graphics with Computer Vision. Springer Verlag.
Frazer, J.H (1995), An Evolutionary Architecture. London:
Architectural Association Publications.
Frazer, J.H. and Frazer, J.M. (1996), The Evolutionary
Model of Design. In (Eds. Asanowicz, A. and Jakimowicz, A.) Approaches to
Computer Aided Architectural Composition, Technical University of
Bialystok, pp. 105-117.
Frazer, J.H. (2002),
Creative design and the generative evolutionary paradigm. In (Eds. Bentley, P.
and Corne, D.) Creative
Evolutionary Systems, San Francisco:
Morgan Kaufmann Publishers, pp. 253-257.
FRAZER, J.H. and JANSSEN, P. (2003) Generative
and Evolutionary Models for Design, Communication & Cognition Vol
36. Nr 3&4.
Gould, S. J. (2000), Wonderful Life: The burgess shale
and the nature of history. London: Vintage.
Graham, P.C., Frazer, J. H. and Hull, M. E. C. (1993), The Application
of Genetic Algorithms to Design Problems with Ill-defined or Conflicting
Criteria. Conference on Values and
(In)Variants, Amsterdam ((1995), Systematica, 10, pp. 61 –
76).
Holland, J. H. (1975), Adaptation
in Natural and Artificial Systems, University of Michigan Press.
Jaanusson, V. (1981),
Functional thresholds in evolutionary progress. Lethaia, 14, pp.
251-260.
Janssen, P., Frazer, J. H. and Tang, M. X. (2003a), Evolution Aided
Architectural Design: An Internet based evolutionary design system. Proceedings
of the 9th EuropIA International Conference, (to appear).
Janssen, P., Frazer, J. H. and Tang, M. X. (2003b), Evolutionary
Design Exploration Systems. Proceedings of the International Conference on
Construction Information Technology (INCITE 2004).
Runnegar, B. (1987), Rates
and modes of evolution in the Mollusca. In (Eds. Campbell, K.S.W. and Day,
M.F.) Rates of evolution, London: Allen and Unwin, pp. 39-60.
Simms, K. (1991),
Artificial Evolution for Computer Graphics. Computer Graphics, 25(4), pp. 319 – 28.