The video above is the final presentation of my second term programming project at the Msc Adaptive Architecture and Computation at the Bartlett, UCL. Underneath you can find a detailed description of the software as well as some photos of phenotypes fabricated for the 2nd term final exhibition which took place at Bartlett’s Wates House on Monday.

You can also take a look at previous posts about this project here and here, although I’m afraid that the explanations there will overlap with the one underneath.

external embryogeny

External embryogeny is a type of GA in which fitness function is pre-defined. In external embryogeny only the populations of solutions are subjects to evolution. It’s opposite is an implicit embryology GA – in which fitness objective evolves together with populations of solutions.

In the following month I’m going to write a paper about “breeding objects” software.
The research question which I hope to tackle in it is whether it is possible to harness an explicit embryology GA as a tool providing design solutions for both well-defined and ill-defined design problems. Natural and artificial selection is incorporated into the software in order to enable such possibility.

ill-defined and well-defined design problems

Well-defined design problem is a quantifiable problem. Budget constraint can be an example of such problem. Performance of the objects – for example the the amount of exposure to the light which they receive might be another example of it.

Ill-defined design problem in a problem which is not quantifiable. Aesthetics is a classic example. There has been many interesting attempts to research general patterns in aesthetic choices during design process, but, to date, every attempt to extract a general rule from the findings failed to some extend. The incorporation of natural and artificial selection into a GA might makes generalization obsolete.

phenotype generative procedure

The advantage of GA is that it can accommodate quite complex generative procedures. The final generative procedure is the 4th one which I was experimenting with. This proves that different procedures can be plugged to the same GA as long as their fitness objective is similar (this is explained in natural selection paragraph). The parameters of final procedure are the size of the object, resolution of the point grid as well as grid’s distortions caused by set of detractors and attractors. Finally range and strength of every attractor and detractor is also a parameter.

natural selection

Natural selection provides an answer to well-defined problems. In other words it is an optimization tool. In case of “breeding objects” software the problem (the fitness objective) is the cost of the object. Since the fabrication method is 2d laser cutting – the cost of the object consists of length of all of the cuts (which determines laser cutting time) as well as the area of the material.

During the natural selection step the software itself determines which individuals it should breed in order to create next generation. The choice is made based on the fitnesses of all of the individuals (the difference between their cost and the fitness objective).

The population quite quickly converges to a fit design solution. However if the process of natural selection continues the it is able to converge into another design solution. This solution is equally fit as the previous one (it’s cost is roughly the same) but it is very different visually.

artificial selection

The convergence from one fit design solution into another during natural selection proves that aesthetics is an ill-defined problem. This is where artificial selection comes in.

During the artificial selection step the user makes the choice about the individuals which become parents of next generation. User can choose just one individual or the whole population as the ones to breed.

The user is informed about the fitnesses of all of the individuals. Their choice can but does not have to be determined by this information. Additional information during this step is provided by 3 viewports enabled in the software: perspective, elevation and shading ratio (the amount of light which the surface underneath every individual receives).

natural & artificial

If the user will deliberately choose individuals with low fitness he or she is likely to loose the achievements of natural selection. This means that the next population can consist of less fit individuals than the previous one. It is not clear however, that it is always the case. What I’m actually going to investigate in the paper is whether that depends on some pre-determined parameters of both natural and artificial selection steps. Those parameters might for example be the percentage of mutation of each individual’s genotype or a range of fit individuals from which parents of next generation are chosen during natural selection.