Text Overview | Examples of Mating & Mutation | Technical Details | Future Directions
In Nature, evolution proceeds by a Darwinian cycle of reproduction, random mutation, and survival of the fittest adult organisms through competition and cooperation to reproduce again. Over the eons, natural selection produces the evolution of diverse species adapted to their environments and each other. I selectively breed my images in a tradition inspired by evolutionary art pioneer Karl Sims. From a library of images dating back to a primordial soup of virtual DNA I began constructing in 1993, I initiate a run by creating a population of around 100 images on a large computer screen. I examine each image and assign it an aesthetic fitness score, then command the population to spawn. Reproduction is accomplished by sexual mixing of virtual genes mostly from the fittest parents, accompanied by occasional random mutation, while particularly fit individuals survive intact into the next generation. A mosaic of new images then fills the screen and the cycle is repeated. An image is born when its genetic structure is expressed as millions of colored pixels filling a frame, as an animal or plant is born when its egg or seed grows into millions of cells each containing an identical copy of its DNA, or genetic instructions. Once some appealing images have evolved, I terminate evolution (a mass extinction except for certain genes preserved in digital amber) and begin a lengthy process of fine-tuning the color and region depicted. Most images saved during the genetic run do not survive this further selection during post-production. Finally I prepare an image for production as an IRIS print or for filming on a film recorder.
Here are some of the ancestors of In the Beginning, the image on my homepage. It would take hours to download even a portion of the actual run to see how the process really works, but I hope to give a slight flavor of it with these examples. Where "+(sex)" is used, there were two parent images. When "->(mutation)" occurs, there is only one `parent'. The following two parents on the left were themselves seeded into this run - their own histories go back thousands of generations, hundreds of thousands of individual ancestors during the period 1993-1996. Note that sex in evolutionary art is not at all like `morphing' in image processing or computer animation; it is not always apparent to the human observer how the offspring of a particular union might look. |
|
2 of 12 unique ancestors in 8th-great-grandparents' generation producing 1 of 13 unique ancestors in 7th: |
|
|
| In the next case the appearance of the offspring is somewhat more apparent. |
|
2 of 13 ancestors in 6th-great-grandparents' generation yield 1 of 12 in 5th: |
|
|
|
Next is our first mutant: an individual incurs a slight mutation resulting in issue which will, ultimately, contribute to the appearance of In the Beginning: |
|
1 of 13 ancestors in 6th-great-grandparents' generation produces 1 of 12 in 5th: |
|
|
|
2 of 13 ancestors in 6th-great-grandparents' generation produce 1 of 12 in 5th: |
|
|
|
|
|
1 of 12 ancestors in 5th-great-grandparents' generation produce 1 of 16 in 4th: |
|
|
| 2 of 12 ancestors in 5th-great-grandparents' generation yield 1 of 16 in 4th: |
|
|
|
|
|
2 of 10 ancestors in 3rd-great-grandparents' generation produce 1 of 7 in 2nd: |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
My current system employs a tree-structured genome, where each pixel is the result of the same evolved program operating in a different region of the plane, which for some functions is the real plane, for others the complex plane. The last image depicted above (In the Beginning) had a total of 427 `nucleotides' or tree nodes:
In addition, each run requires dozens of parameters determining initial conditions, some of which can be randomized and set automatically at the beginning of a run, others must be set manually. Certain regions of phase space (initial conditions) seem to work well in deterministic chaotic functions embedded in a genetic art system, others do not, suggesting basins of aesthetic attraction in evolutionary art hyperspace. It takes so long to evolve worthy images from a completely random primordial soup - hundreds of generations - that for the last two years I have been seeding my new runs with preserved genes from past favorites. Over the hundreds and thousands of subsequent generations, the genetic library itself becomes the most important and valuable part of an evolutionary art system, more even than the extensive software libraries themselves.
International Standard for Cooperative (& Competitive) Evolutionary Art? Does anyone have the time to design an open standard for the interchange of genetic art elements? Musicians are passing around, mutating and replicating MIDI sequences that behave as musical memes. The same could be done with images. The grapevine has it that Karl Sims may be developing a commercial 2D genetic art language. Once capable software for evolving images on home computers is readily available, an open standard would enable similar cooperative interchange of genes producing components of images. For a survey of what software is currently available see Andrew Rowbottom's site. My own code is non-portable and highly optimized for Silicon Graphics computers. This suggestion is related to Tom Ray's proposal for a `digital nature preserve' - a portion of the global internet devoted to open-ended Tierra evolution. If someone gets open-ended genetic art evolution up and running coupled with this international standard, it could run in tandem with people registering their fitness votes, as is currently done in Absolut Kelly. Could Chris Langton's SWARM system (click here for a link to the SWARM website) be used as a common vehicle for generating evolutionary art? Open-ended genetic art evolution:co-evolving system of pattern generators and recognizers: I had hoped to begin implementing my design for a co-evolving system of image generator and image commentator organisms (as neural networks themselves evolved by selection) during 1996, but time keeps running away from me. Jon Schull suggested a similar idea to me at Artificial Life III, tracing its origins to an idea of Richard Dawkins', which would involve open-ended image evolution with the fitness function to be determined by real bees - you want to see if images of actual flowers would evolve. Such a system would require hardware interfaces and funding, but an internal system such as I have in mind is within reach of the individual investigator. An evolutionary artist would bootstrap the process by guiding the evolution of aesthetically pleasing images for some number of generations, say 20. Then you fire up the population of evolving neural networks, and have them cycle for however many generations it takes to come sufficiently close to your aesthetic fitness selection as they `look' at all the images in the populations of the 20 bootstrap generations. Once that happens, you turn over the aesthetic fitness selection for the 21st image generation to the "newly confident" neural network organisms - and the artist steps back away from the hands-on control by one level. Now you can let the populations of image generators and commentators co-evolve automatically for zillions of generations. I maintain that no one knows what this process might reveal, or how far it might go. In principle, by analogy to the supremely powerful metaphor of the co-evolution of flowering plants and pollinating insects, it could lead to the evolution of quite powerful, perhaps universal shapes or forms. Coupled with a realtime source of random numbers from Nature [rather than a deterministic pseudorandom number generator], might it be our first chance at an "archetype camera"? The artist could experiment with creation itself by varying subsequent involvement. Totally hands-off: you have a Free Will universe gifted with only its initial state inoculation - see where it goes. Or reach in and tweak it from time to time to keep the evolution appealing to a human aesthetic. |
+(sex)
= 
+(sex)
=
->(mutation)
=
+(sex)
=
+(sex)
+(sex)
+(sex)
