Abstract

By evolving design concepts dynamically using computer models in conjunction with genetic algorithms the computer in effect compresses evolutionary space and time by allowing large numbers of iterations of small variations of a model to be tested and evaluated in a simulated environment. The use of genetic algorithms is now a well established technique in many technical and engineering design applications. This paper describes an adaptation of the technique for three dimensional design applications.

The first example is from yacht design where data on displacement, wetted area, block coefficients and other factors giving an indication of likely performance are considered along side ergonomic data derived from the sections, and aesthetic and intuitive judgements derived from considering graphic displays of the boat lines. This allows both illdefined and conflicting criteria - a common situation in design - to be considered using the same technique.

The second example describes an experiment with trying to cooperatively evolve

a dynamic environment on a global scale through the medium of the Internet. The technique relies on genetically coding design concepts and evaluating a pool of biodiverse genetic material contributed by global participants.

Design by Evolution

Evolving design concepts by mutating computer models in a simulated environment is now a well established technique in fields as diverse as aeronautics, yacht design, architecture, textile design, fine art and music^1,2. The technique used in each of these fields involves a 'genetic algorithm' which requires the key parameters of a design idea to be encoded in a 'code script' which is analogous to the code script of life encoded in DNA³. In a similar manner to evolution in nature, this code script is subject to crossover and mutation and the resulting forms developed from the coded instructions are subjected to some form of selection⁴. Clearly the actual encoding of the instructions varies from discipline to discipline but it is in the selection process that significant differences occur. In those design fields where it is intended that the evolutionary approach should yield an increase in functional efficiency, then the genetic codes which yield the most efficient results are selected. This is seen as being analogous to natural selection. Where the selection criteria are less easily quantified or the criteria are more concerned with aesthetics or personal choice, then a technique of visual, judgmental or intuitive selection can be used. This technique is more analogous with artificial selection, which is the process used for breeding race horses, greyhounds, pigeons, cattle or any domesticated animal. A breeder uses their skill, experience and judgement to decide which animals to mate or select for breeding stock, and similarly a designer can use their experience and judgement to select genetic variants for further experimental development. This can be done by examining a 'population' of genetic variants which may be presented graphically and a selection made on the basis of visual inspection (this is the technique used in Richard Dawkins Biomorph program⁵, and several fine art and computer graphics derivatives^6,7). This visual display can be supported by further statistical and analytic information as the authors have done previously with yacht design where data on displacement, wetted area, block coefficients and other factors giving an indication of likely performance are considered along side ergonomic data derived from the sections, and aesthetic and intuitive judgements derived from considering graphic displays of the boat lines. This allows both ill-defined and conflicting criteria - a common situation in design - to be considered using the same technique⁸.

How an evolutionary model works

The evolutionary model requires that a design concept is described in a genetic code⁹. The code is then mutated and developed in a computer program into a series of models in response to a simulated environment. The models are then evaluated in that simulated environment and the code of successful models is selected. The selected code is then used to reiterate the cycle until a particular stage of development is selected for prototyping in the real world.

In order to create a genetic description it is first necessary to develop a design concept in a generative manner capable of being expressed in an a variety of forms in response to different environments. This is a manner in which many designers already work in the sense that they have a personal set of strategies which they adapt to particular design circumstances. This strategy is often very pronounced and consistent to the point where a designer's work is instantly recognised.

EXAMPLE I - Yacht design

The design of yacht hulls is a task for which there are objective measures of efficiency including stability, centre of buoyancy, wetted surface area, prismatic coefficient, block coefficient and so on. Taken together these measures can give some indication of likely hull performance and potential speed¹⁰. There are also ergonomic criteria, or measurements relating to racing rules. All these criteria can be awarded a weighting depending on the use of the hull (racing, day sailing cruising etc) and incorporated into an overall assessment of a fitness score.

In an interactive evolutionary model used with a group of second year design students at Ulster in 1992 and developed by one of my doctorate students, Peter Graham¹¹, a simplified version of his genetic algorithm based program allowed a hull to be primarily assessed for potential speed on the basis of the prismatic coefficient (which is a measure of the fullness of the hull), and ergonomic and other considerations were assessed by the design team by studying plotouts of hull shapes proposed by the program. Thus two scores, one from the analysis program indicating potential performance and one from the team on other less well specified functional and aesthetic criteria, were combined to control the development of the generative program.

The program generates a randomized initial population of hulls for which the prismatic coefficients are calculated. The design team were then presented with the range of hulls to score on the basis of their criteria. The score from the designers and the degree of divergence form a target prismatic coefficient are then subject to a simplified Pareto optimality test. This consists of testing the degree of domination of each hull within the population, that is the extent to which it is bettered by others in terms of both criteria, and this is taken as a measure of fitness. On the basis of these fitness scores parents are selected, preference being given to those with the highest fitness scores using biased random selection. A new population is then produced and the process repeated. The resulting hulls demonstrated the feasibility of this technique for combining objective performance with less clearly defined and potentially conflicting criteria.

'Natural selection' can be used to optimise the quantifiable criteria whilst the process can be periodically interrupted to make a more intuitive decision or 'artificial selection' (or an intuitive development can be handed over for technical assessment and fine tuning). This procedure closely relates to the actual working methods of top yacht designers who use sophisticated computer techniques in conjunction with traditional skills for establishing initial boat hull lines¹².

So far this implies a form of convergent evolution where the object is to produce one optimised solution, but this need not necessarily be the case, and the technique can equally be used to explore divergent evolution with the intention of encouraging a range of widely differing alternative solutions. This may be with the object of producing a range of choices for a client, or perhaps to broaden the search area, or simply because we have run out of new ideas. Again both divergent and convergent evolutionary techniques may be combined so that first the search space is widened with a divergent approach, and then a range of competing ideas are convergently and competitively evolved.

EXAMPLE II - The environmental model

This was a collaborative research project involving both the University of Ulster and the Architectural Association. In the first stages the team at the Architectural Association in London had developed a theoretical model of an evolvable environment. For the example described here, Manit Rastogi of the Diploma Unit 11 at the Architectural Association developed a working interactive computer version and teamed up with Peter Graham from Ulster to produce a dynamically evolving model which was exhibited in London and launched at a site on the Internet¹³.

The model is organised using a multiple hierarchical approach and a datastructure which is recursively self similar¹⁴. The simulated environment in which evaluation takes place is modelled in exactly the same terms as the evolving structures. The environment and the structure not only evolve in the same dataspace, but can co-evolve, and competitive structures can also evolve in the same space. Environment in this case includes user response and is modelled with virtual societies. The environment has a significant effect on the development of the concept which is described using a genetic design language. Genetic algorithms are used to perform the selection and normal crossover and mutation are used to breed the populations.

The model consists of an endless array of data points which collectively constitute a dataspace. Each point in the dataspace is intelligent in the sense that it knows where it is and why it is there and it has a clear awareness of the spatial relationship of its neighbours. The laws of symmetry and symmetry breaking are used to control the development of the model from the genetic code. Information flow through the model takes the form of logic fields. Externalization of this datastructure is process driven by modelling the process of form generation rather than the forms themselves.

The model is based on the sequential evolution of a family of cellular structures in an environment. Each cellular structure begins development from a single cell inheriting genetic information from its ancestors and from a central gene pool. Each cell in a cellular structure contains the same chromosomes which make up the genetic code. The cells divide and multiply based on the genetic code script and the environment with each new cell containing the same genetic information. The development process of each member of the family consists of three parts - cellular growth, materialization and the genetic search landscape. A genetic algorithm is used to ensure that future generations of the model learn from the previous ones as well as provide for biodiversity during the evolutionary process.

The datastructure of the model is based on a universal state space or isospatial model where each cell in the world has a maximum of 12 equidistant neighbours and can exist in one of 4096 states, the state of a cell being determined by the number and spatial arrangement of it's neighbours.

The local environment of a cell in the world can thus be coded in a 12 bit binary string. The growth and development of the cellular structure is controlled by chromosomes.

Chromosomes are generated by either being sent in by any remote user, an active site or as a function of selection, crossover and mutation within cellular activity and are maintained in a main chromosomal pool. The physical environment determines which part of the main chromosome pool becomes dominant. The local environment of each cell then determines which part of the genetic code is switched on. The cell then multiplies and divides in accordance with that genetic code.

As cellular division takes place, unstable cells are generated. In the next generation this leftover material creates a space of exclusion within the cellular space. This space of exclusion interacts with the physical environment to create a materialization of the model. Boundary layers are identified in the unstable cells as part of their state information and an optimized surface is generated to skin the structure. This material continues to exist throughout the evolution of the model and will initially affect the cellular growth of future generations.

The selection criteria in the model is not defined but is an emergent property of the evolution of the model itself. A genetic search landscape is generated for each member graphically representing the evolving selection criteria within the model based on the relationship between the chromosomes, cellular structure and the environment over time. Form, or the logic of form, emerges as a result of travelling through this search space.

Once chromosomal stability has been achieved, the parent cellular activity is terminated. The final cellular structure, the materialization and the genetic search space are posted out. A daughter cellular activity is then initiated from a single cell. The fittest chromosomes from the parent generation are bred using selection, crossover and mutation and combined with the new list of dominant chromosomes from the main chromosome pool to form a new chromosomes set for the daughter generation. The development process is then repeated for the daughter generation.

Conclusions

The model was launched on Internet on 25th January 1995 and in the first two weeks it evolved four family members based on the chromosomes received and those bred internally, each member achieving chromosomal stability in about 120 generations. Though it is impossible to predict the nature of the model yet, or its evolving internal logic, there seems to be a pattern emerging towards its selective and hence, evolutionary process.

The next step is to recode the model so that it can evolve on any computer platform, eventually making it completely autonomous on the Internet. The model could then evolve indefinitely by allowing the possibility to replicate itself onto any host computer.

Overall we have had a great deal of feedback and positive response from those who have interacted with our dynamic models and we are moving rapidly to producing more general purpose design tools based on these techniques.

References

1. Frazer, J.H. An Evolutionary Architecture, Architectural Association, 1995, 127pp.

2. Forsyth, R.(ed), Machine Learning, Chapman and Hall, 1989

3. Goldberg, D., Genetic Algorithms in Search, Optimization and Machine Learning, Addison_Wesley, 1989.

4. Holland, J., Adaptation in Natural and Artificial Systems, Univ. Michigan, 1975.

5. Dawkins, R., The Blind Watchmaker, Longman, 1986.

6. Todd, S. and Latham, W., Evolutionary Art and Computers, Academic Press, 1992.

7. Simms, K., Artificial Evolution for Computer Graphics, Computer Graphics, Vol 25, 1991.

8. Graham P.C., Frazer, J.H., Hull M.E.C The Application of Genetic Algorithms to design problems with ill_defined or conflicting criteria, Conference on Values and (In)Variants Amsterdam 1993, Systemica, Vol 10, 1995, pp61_76.

9. Frazer, J.H. The Genetic Language of Design in Textiles and New Technology: 2010 ed S Braddock & M O'Mahony Artemis London 1994, pp77_79

10. Hammitt, A., Technical Yacht Design, Adlard Coles, 1975.

11. Graham, P., Evolutionary and Rule_Based Techniques in Computer_Aided Design, Doctorate Thesis, University of Ulster, 1995

12. Holland, R., Splendour Under Sail, Shoreline, 1988.

13. Frazer, J.H., Graham P.C., Rastogi M. Biodiversity in Design via Internet, Proceedings of Conference Digital Creativity, Brighton April 1995, pp97-106.

14. Frazer, J.H. Datastructures for rule_based and genetic design, Visual Computing _ Integrating Computer Graphics with Computer Vision, Springer_Verlag Tokyo, June 1992, pp 731_744.

For more SYNDICATE SPEAKERS

For other SYNDICATES

Cyberbridge-4D