Biota.org

BIOTA PUBLICATIONS

The Artistic Spectrum of ALife by Sue Wilcox Steven Rooke uses evolutionary techniques to produce aesthetically pleasing abstract images without having a conscious plan how they should look. I asked him where he comes on the spectrum of artists using ALife? He unequivocally places himself in the tradition of Karl Sims 1991 work, evolving 2D images, as opposed to Sims 1994 where he got into evolving form and function in 3D organisms. The main point of this stage of Sim’s work was that he evolved images without designing them. Steven describes his process of arriving at images as being producing pictures without knowing what he wants to put there. Things would emerge that were coming from his brain without his knowing what was to go where. Without his having a plan and without his even considering himself to be an artist. He said "I don’t think of myself as an artist. When I started out I was a geologist. Then I became a systems programmer and an astronomer. I don’t know art-speak - my eyes glaze over - and I seem like an idiot." However with the encouragement of Ralph Abraham, one of the architects of chaos theory, Steven quit his job and within two weeks saw all the imagined obstacles to his becoming a computer artist melt away. He set up with an SGI machine and began writing code feverishly. At the beginning as he describes it "it was a total primordial soup - there was almost nothing to work with". The Process: To start with he needs to prepare a space to hold a ten by ten grid of 100 pixel square images - the organisms of a sort that he intends to evolve. Then he creates a ‘tree structured genome’ which is filled from components he gives it. (This is the same architecture described by Karl Sims in his 1991 SIGGRAPH paper.) The tree structure is the genotype. The phenotype is expressed when you run the program. And it fills in all the pixel locations allocated to it on your screen. *resolution? Its the same program that runs to produce each pixel of these images, the only thing that’s different is the x and y coordinates. "Each organism has a different instance of a tree. One might be a small tree, another might be a big bushy tree and they’d have different functions in them but they’re all trees." And there are a hundred different trees displayed on screen at the end of a complete run of one generation of the evolution process. "Each picture is one organism." The tree is an upside down mass of constants and equations and functions. If you wrote it out as computer code it would read as a series of IF, and ELSE statements which would return a number for a color to appear in one pixel of the computer screen. But in that form it could not evolve. So the tree is a series of branching rearrangeable possibilities for calculations that work with the location of a point’s x and y coordinates. The relationship of the colors to the coordinates via the intermediary of the tree program produces the image. Equations and functions take the value of the pixel x and y coordinates and process them through the many formulae in the tree. The functions and equations constituting the branches of the tree are anything you can make up; a mixture of standard trigonometry and arithmetic functions plus subroutines to generate more complex objects e.g. fractal non-linear equations. The branching stops when you get to a leaf - a constant like pi or a number. "You throw your whole math library at it - arc, tangent and square root all the functions you can think of" says Steven. "By the time you’ve got done you’ve got this huge tree with bunches of X’s and Y’s all over the place. The leaves are all sorts of constants, like pi that you know are going to be useful. And then for the functions, the branches of the tree, is anything you can make up. So you put in the obvious standard things: plus and minus, times, and basic arithmetic, simple functions. And then I wrote a number of subroutines that function the same way but generate more complex objects. Such as the Mandelbrot set." These dynamic functions add visual complexity to the final images as they cause the program to move between normal Cartesian geometry to the phase space where the strange attractors of chaos theory function. It’s like tapping into another dimension. Steve describes it as: "What’s coming up this branch of the tree may be coming up from way down there. So what the program thinks is x and y could be all sorts of things. Once it gets in here [phase space] it goes through this attractor thing for a number of iterations So what comes out at the end of it is not something spatial it’s a dynamical system measure of how close something was in coordinate space to an attractor. A phase space attractor. So you’re going from spatial space, Cartesian type space to phase space. And that’s important. What the genetics does is totally scramble all that up. The genetics constructs these trees that are constantly flipping back and forth from Cartesian to phase space. By the time all this gets done you get a number out the top. And that number ties to a color map and that becomes the color of the pixel." It’s an astonishingly elaborate way to derive a color for a pixel, and takes considerable number crunching power from a computer to accomplish. But the results are impressive. Judgment: The artist gives aesthetic fitness scores between zero and nine "small numbers so I can do it fast" to the hundred organisms generated by the program. "The granularity is pretty large - there’s only ten different scores I can give to these organisms. That tends to get mushed out by the probabilistic nature of the selection procedure for the next generation. At the beginning they all have a fitness score of zero. In the early primordial soup rounds I assigned a score higher than zero to ones that had at least some interest to them." Then probabilistically the ones with the highest scores get to be cloned or breed together. This must make Steven one of the most practiced aestheticians on the planet - constantly assigning numerical artistic value to images. Once a score has been assigned to every member of a population its time to produce a daughter population. "I’m not sure that I’m entirely claiming that its art - it might be something entirely different. It’s an exploration of evolution. That’s why I do it. The tools I use, instead of paint and easels are compilers and datastructures. So I think of myself as a programmer - a designer of complex systems software. And I just pick the images that I like and give them a score from zero to nine. However, if you just bred the best and brightest in every generation you’d get there a little faster - you’d get something interesting to look at on the screen but then you’d get stuck. Everything would start looking the same. You’d have a very restricted gene pool. So you’ve got to preserve some runts. If you get a strain of potatoes or wheat or rice that is extremely productive and you put all your money into that one - things will be great for a while, but then some virus will come along and wipe out your crop and you’ll need to go back out to the wild and get some old ancestral rice. You need diversity. You need genes that express themselves in a way that is not advantageous in the current generation. But if you don’t have them there their descendants are going to pay for it. " Steven proceeds to copy the best organisms and to change or ‘breed’ the others. To produce the effects of mutation and sexual selection the program simulates these processes. He says "I played with different ratios of sexual cross-over, mutation, and the cloning of the most fit individuals from the previous generation." Within the genome tree it is possible to swap sub-branches - this is used in the cross-over process. Steven finds a proportion of 60% of the population used for cross-over experiments is good, with 20% of the population used for cloning, and 20% subjected to mutation. Out of his hundred organisms he copies or clones the 20 fittest individuals - these are the ones he likes the look of most. Then he selects 30 pairs of organisms to swap sub-trees using a random number of swaps - this simulates the sexual exchange of genetic material. "In biological meiotic sex you have two strands of DNA - sexual reproduction means that when you’re creating a sperm cell you have half this many nucleotides going into a single strand of DNA. With trees its a little different because they’re not linear. You just take a sub-tree and pluck it out then you pluck a different node out from the other tree and swap the sub-trees." In fact with the trees this process is repeated a random number of times. The last 20 organisms get mutated, they only have one parent - they are essentially changed within themselves - they get one function flipped at random. Steven finds that most mutations are not useful. He describes the results compared to the parent: "Often its unrecognizable, sometimes its similar, sometimes just a little bit different, but most mutations are bad, as they are in nature" i.e. he doesn’t like the look of the result "but occasionally you get a good one. It’s probabilistic breeding. The ones that have higher scores probabilistically will tend to breed first. ". However he believes that sex produces the best results. "And now you have a daughter population displayed on the screen. And because you selected for interest the first time you now have a population with a lot more interest." As you can see from the weighting he gives to sex Steven Rooke believes the best results come from cross-overs. "Karl Sims does it a different way around: he had a much smaller group of organisms (16) and he mostly used mutation rather than cross-over. But I think the reason I got good results was that I was more of a fanatic than Karl was, I spent more time plugging away at it plus I had larger populations." However it takes more than just running the program - its essential to understand how a gene pool works: "If just bred you’d shrink the gene pool too much. Probabilistic breeding is important. The runts are crucially important. Plant breeders know this." The whole process is an exploration of evolution - there is none of the usual art speak encountered when artists talk about their work. There is no art theory here - the fascination is with the forms produced and the software process that creates them. Although the end results of this evolutionary process can be beautiful there is also an attraction in watching the process itself. How long does the process of evolving and selecting go on for? With an initial setup of 100 organisms, 40 generations takes around 2 days depending on the size of the genome. "I would always work with as large a genome as I could manage and an amount of time I could manage without going addled." Steven lets the computer run and calculate images (in 8 bit 256 colors) then works at selecting and grading images for 10 to14 hours a day for 2 or 3 days. "At that point I can’t take any more. It’s mass extinction time. I hit the Escape key." Which stops the next iteration from happening. Particular individuals are selected as seeds for future runs, he saved 10 or 15 from each batch in the early days. They are preserved as the complete genome: C source code, a compiled program that can be run to generate the image again and again. "I wrote a reverse compiler that translated the genome back into C source code. Then the evolutionary program writes out an entire C source code file which then gets compiled. (This is in contrast to the trees which are an interpreted way of expressing the same idea except that they allow the evolutionary maneuverings to take place.) So at the point of mass extinctions I have 15 to 20 compiled programs and images that are a little bit bigger to look at - about 200 pixels across. The next stage is to start working with those images." Post evolution is cropping and color: Only a part of an image may end up being printed. Steven uses a special program that allows him to enlarge the image. The original image is usually cropped and one area zoomed in on to produce a particularly striking result. Then Steven tries out different color maps than the one used for the evolution runs. "I always tweak color last. A good composition comes first." Steven uses some programs he’s written to produce more complex color maps than he feels he could come up with alone. "Because I want to evolve these things in color I start with a color map. The color maps themselves are co-evolved. The 10 x 10 set of organisms on a screen, during evolution are all in the same color map." He does evolution in 8 bit color because its faster, then moves across to using 24 bit 16 million color for the final images to enrich and intensify the colors. But usually finds that the color map an image evolved with is the one that remains the most pleasing. "Almost always the color map that you evolved the images under is the one that works the best for final production. The color is involved in composition. Though maybe a tweak is a little better." Making the transition from 8 bit to 24 bit color takes some adjustment. "When I move into 24 bit color space I start the color maps as vectors (RGB) - vector tie points. So, looking at the horizontal axis as the range of the number at the top of the evolved program, which I scaled from -1 to 1, color intensity is the vertical axis. You have a bunch of line segments between the tie points, making up three wiggly lines in red, green, and blue. During evolution that horizontal axis from -1 to 1 maps to 256 possible values to represent different intensities for each color. But after that in post production I stretch that 256 out to 65k (16 bit color for each of red, green, blue) and that means the precise angle at which the line segments cross each other causes subtle new colors to emerge from in between the cracks you couldn't see when the range was only 256. Each one is really a 65k (16 bit) color map, but you need a 24 color monitor to cover the range of different color maps. You do see a big difference between the 8 bit and 24 color modes - in 8 bit there are a lot more contours between the colors. And when you go to an Iris print there's a lot more pixels than you see on the screen, so get more benefit from the stretched-out color map. I haven't seen colormap contours on a big IRIS print with a 65k colormap, but it would be easy to stretch it out further if you did." Then he makes a 35mm slide from the image, " in the early days I used to do that off the screen using a camera and a tripod" using a 1280 x 1024 high resolution SGI screen. Since ‘96 Steven uses a film recorder to produce his high definition images on film when he needs them in slide form. The film recorder has a very tiny computer screen inside it - capturing one scan line at a time in three separate passes with red, green and blue filters. He wrote software that sends the right instructions to the film recorder to capture the images. This captures serious detail: for a slide run he generates each image at 4800 x 3200 pixels which is 45 MB image uncompressed on disk. The film recorder has a resolution of 4800 x 3200 pixels to produce a photographic slide. But for the highest quality images on paper Steven prefers to print onto an Iris printer direct from the computer file. Getting the colors to reproduce accurately is a notoriously difficult process. He saves the images as TIF files and uses the genome in the image header to make sure he doesn’t lose it. Printing: He uses Bishop St. Press in Kihei on Maui with Piccadilly Press as the distributor. Conversion to CMYK is a difficult process; he prefers LivePicture's RGB-CMYK converter. He uses ICC profiles to calibrate his monitor to the printer's particular ink & paper combination - for this task he had to use a Mac because LivePicture was only available on a Mac. Now its WYSIWYG. He does his own color separation and burns a CD (>400MB) in the correct file format for the printer. Print is on high quality Somerset velvet archival watercolor paper. The next run: Steven finds that seeding a new run with some of the best images from previous runs speeds up the process of arriving at aesthetically pleasing results. Starting from a primordial soup each time just takes too long. It took him three years of experiments to evolve his present stock of ‘seed’ organisms. He says that the richness of an image comes not from the size of the genome but from its internal structure - which comes only from lots of breeding and from the use of a large genetic library of seeds. Seeding the primordial soup hurries things along. New work: In 1998 he added Genetic Fractals, which employ the genetic search process to construct brand new, arbitrarily complex fractal equations that would take a ridiculous amount of time to make up and try out by hand. Fractals are images derived from algorithms based on chaos theory using non-linear equations. Sometimes the results are a sort of landscape which remind the viewer of scenery rather than the completely abstract look of the earlier work. Collaborators in Tucson project slides onto dancers as performance art pieces. He has implemented the Genetic Cross Dissolve introduced by Karl Sims in 1991, using both his older genome and the new embedded genetic fractals to produce animation. The cross dissolve produces a lot of animation "for free", as it traverses hundreds or thousands of previously unseen genetic relatives as it goes between genetic `keyframes'. Coming next: A new genome using directed graphs rather than tree structures so the branches can come back together instead of endlessly dividing. This circuitry introduces the possibility of experimenting with simulated metabolism and running several iterations of a program to gain cumulative effects. With this more complex genome Rooke believes he can co-evolve separate populations of organisms that look at the evolving images and make judgments about them, kind of like the co-evolution of flowering plants and pollinating insects. "What I'd really like to see would be the evolution of aesthetic parasites and corresponding aesthetic immune systems. Then things should get really interesting." But small isn’t the only way to be beautiful. Rooke has work on a larger scale using over a billion pixels in one image. This 11 foot by 7 foot mosaic of 9 full-sized IRIS prints in a 3x3 grid, containing 1.1 billion pixels, is currently looking for a home. He’d like to try even larger works done on sixty foot walls but using mega resolution so walking up to a piece gives the deep zoom effect you can see on a computer generated image. "It would do something neurological to you, walking up to it from across the plaza". Look out Wyland whale walls may be replaced by psychedelic ones.