Karl Katzke

Kottke’s feed linked today to Smashing Telly’s post of a youtube video in which an architect experiments with allowing a computer to generate solutions to a hotel design. The conclusion is that the computer’s designs, even though they’re using a genetic algorithm to ‘breed’ better designs as time goes on, that the “best” solutions posed by the computer don’t make any sense in a human world. Therefore, the computer is useless as a tool without the design inspiration that’s brought to the drafting table by the simple addition of human experience.

I take issue with the conclusion that the architect reaches. I’m not sure if the limitation is the 3dmax scripting language itself or if it’s the author’s programming skills — but no “genetic” algorithm is going to evolve a perfect design unless it’s given continuous feedback from an evaluation agent. The evaluation agent in the case of a draftsman at his table is the experience, including the human desires, that the draftsman possesses. Without digitizing twenty to thirty years of human experience and however many number of years personally experiencing and evaluating architecture, how the bloody hell did the architect in this progress even expect to end up with a workable solution?

From the video, you can get a couple of ideas of the types of criteria that the program was fed. Hallways should be sized so that they dampen acoustic noise. There must be a restaurant that is easily accessed. There are core services that need to be provided to all users of the building and they should be centrally located. There need to be multiple sizes of guest rooms. All guest rooms must be accessed via a hallway. And then there seems to have been a group of parameters fed that lead to solutions that -don’t- make any sense to a human … like insulating the guest rooms from one another. Aatmosphere and an external wall are the best insulators, so obviously a computer program would choose that solution, if given the criteria that a room should be made as quiet as possible.

I’ve been doing a lot of work with genetic and simulation algorithms for a computer game that I’m working on… or in other words, my idea of fun. The problem set I’m working with is limited enough (or I’ve made design decisions that have broken it down far enough) that I can provide the program with adequate limits. Anywhere that I can’t provide adequate limits, I’ve arranged to use some form of human evaluation — Amazon’s Mechanical Turk, requests for decisions by players, or a database of previous choices made that the program can “learn” from. Other actors are required to tune and refine the results of a simulation.

People, computers are only as smart as the program they’re fed. You don’t end up with a magic solution (what the video’s creator seems to have desired) by just running a program for long enough. Not going to happen. This is one of the arguments against evolution as the source of human beings… but if you think of fauna as independent beings, you realize that there are other actors, and the methods of evolution that lead to better products are designed such that the actions of the actors get the desired effects.

For instance, let’s say that a variety of plants are competing with one another. The solution so far has meant that the plants bear fruit, which is eaten by animals, and then the seeds are fertilized by being dumped along with the animal’s waste. One plant’s random solution is to make it’s berries sweeter. The second plant’s random solution is to make the seeds more bitter. The third plant’s random solution is to make the berries toxic, which causes the animals to gag and vomit after eating it. The animals that feed on this fruit are runners, and cover a vast distance in a day.

The predictable results of that random variation would mean that the sweeter seeds would be distributed far and wide, and more heavily consumed. The more bitter seeds would force the animals to seek water to quench the bitter taste with, and as a result may be found closer to water sources like streams and ponds. The vomit seed plants would only be found close to one another and would likely grow in recognizable patches — and later in their evolution, may die out as animals realize that “plants that grow in big bunches aren’t good to eat” … and if you think animals aren’t that smart, I’d like to introduce you to my dogs, who have a very acute sense of where “good things to eat” are, and as a result know how to open the dishwasher and refrigerator with their mouths.

Without those fruit-eating animals? All you’d get is more random variation, without the discreet . The plants would drop seeds close to the ‘parent’, and the sweetness of the fruit becomes inconsequential. The selection then comes from who can outgrow the other. Note that there are still agents involved — the other plants. Hotels (in some cases, unfortunately) don’t eat or outgrow one another without being utilized.

The movement towards computer-evolved designs in and of itself is interesting, but to technologists it has been inevitable. Unfortunately, this “experiment” has left out a huge chunk of the process that makes computer-evolved design work, and the conclusion that architecture is impossible without the draftsman’s human hand guiding the process is as fallacious, poorly thought out, and scientifically unsound as humanity’s persistent belief in intelligent design.