Evolving 'portrait painter programs' using Genetic Programming (Darwinian evolution) and a portrait of Darwin.
(this is a work in progress - excuse the crude site and text ---- open your browser page this wide) --------------------------------------------------------|


Source (fitness function image) and an ordered progression of the best of the 33 days (so far) but best is less interesting than most creative family members ..

This is a conceptual piece that creates and evolves a related family of abstract portrait painters. Each Portrait below is created via one evolved computer program. These programs are created and evolve by Darwinian evolutionary techniques (crossover, mutation, survival) using the automatic programming technique called Genetic Programming (see tech and credit details at the bottom of this page). The environment they prosper and have offspring in (or not) is the resemblance fitness function of the most famous Portrait painting of Darwin (1st image above). The 'most fit' (resemble the portrait better than their neighbors) of a population are 'married' together to create 'more successful' offspring. The genes (or function set) were not specifically created for the portrait painting task, but were our genes from what we evolved from any different -- surely not originally created for writing computer programs or making art. But evolve they must.

Can you bring the ghost (creativity) out of the machine using the ghost of Darwin (his namesake techniques and portrait). This is a work in progress piece on things human (the creativity of art, the way life evolved) and things creatively computer - and in the process evolving portraits that are alive (they are not just images but programs with evolved painting strategies that can be reused and married together).

Here is a sample of the last 33 days of evolution ( RUN on 1 PC). The portraits below are in order, starting with the first population and moving in chronological order to the present. I have culled my favorite portraits. These are but a few of the 1000s that slowly evolve towards the Darwin portrait below. With regular Genetic Programming (GP) the end result (“optimization”) is what is important. With this process it is more about the journey ("creativity"), while the overall population gets better at Darwin's resemblance, that is less the point (the Darwin goal is simply the creative spark or more specifically the environment). Since I save the genes of each portrait, I can combine (marry) and re-evolve any in new variants, see HERE . I have both picked portraits that were the best at resemblance as well as "strange uncles" who while less strong at the main goal as thier dominant relative, are still artistically compelling (most below are creative "strange uncles" - as are most artists). Notice as you browse, how as in real evolution, certain strategies remain in the gene pool, eventually combining and mutating with other strategies into new combined strategies. I intersperse the source image of Darwin throughout, so you can compare. Remember the goal is not to recreate the Darwin portrait but to use it (and Darwinian techniques) to evolve abstract portrait painters from the computer primordial ooze. These then are an alive, related family tree (100, 000s generations) of portrait painting programs. - Steve DiPaola (evolve@dipaola.org) - comments? and requests for interbreeding like HERE.

Feb 1 We start evolving portraits -- after a few populations -- we get the color -- with a simple ramped split and then curves ..

At 100s in, mid tones mingle, then a first strategy appears: bands can resemble the strong vertical lighting of Darwin portrait (4,5 below)

this vertical band strategy holds for many runs, but eventually twisting and manipulating the bands ensues ..

this twisting and curving of the initial vertical bands continues, but 2 below the bands thin, allowing more subtleties

Feb. 02 - 20,000 populations in , then combined with imbedded strategies of curves and twists - till this last (5) 'face glow' forms

this glowing face hi light (5 above) is now the dominant strategy; slowly the glow/banding/curves evolve to the first 'heads' shapes

colors and forms erupt, as the quicker early evolution slows down, making slower progress but going through a very colorful stage

Feb. 04 we stay caught on this evolutionary plateau, repeating themes and beginning to use "genetic drift" to slowly find a new niche ...
an interesting irony (which I exploit), is when the system drifts and loses it way (from resemblance) is when I think it is most creative
note how, like real living organisms, the programs reuses and intermingles it best genes (strategies)
-- note the repeated themes:

the stage that began with the 1st image below and went wandering, ended up at the next major dominant strategy - the 2nd image

Feb. 9th: this 2nd image above, with its strong hilight and gradation to skin color, took over the niche, heralding in the soft blobby age

this soft blobby 3d form flourishes with many sub strategies, the right eye above (1) doesn't hold right away but does resurface later
a new group of themes emerge, as we slowly progress and evolve, note the reoccurring forms

Feb 12th: A new addition strategy forms, the soft blob form first picks a dark patch (1 below) and it evolves into the eye area (2, 3, 4, 5)

this "left raccoon eye patch" becomes dominant and combines with other repeated themes, including the return of the right eye

skin texture which evolved in images above, initially above the eye patch, spreads all over ( skin texture? )

stuck in another plateau, all themes dance around, but mostly the texture + negative (& positive) space of the right eye and left eye patch
Feb 18th hardly evolving, the wondering begins and I let them play with texture by increase image complexity (size of allowed program which makes for slower runs because of the additional genes)

Feb 20th Evolving away towards Darwin, 20 days and still going. While they are still not at a realistic resemblance, they do have an abstract center head shape, remember many I have picked are not the best of a run (best resemblance at that time) just ones that show the creative flow and strategies.. The two below are the current best fitness but again the journey is more important the result - these are programs with genes of evolved visual strategies - we always had a final Darin - what we're after is to evolve painters from computer ooze.
Feb 27 - then a more painterly era starting evolving, this first image below, in some form, being the most evolved but with many variants ...

march 5th - below 1st is the current best - showing a primitive beard,skin and eye shadow colors (compared to old best, 1st 3 above)

Note how, even late in the run, even though all painter programs are related, you can still get very a-typical painters, the 2nd and 4th above are examples, diversity can still make for some strange bedfellows.

Still runnung - more updates soon. -- dipaola.org - sdipaola@sfu.ca



Some Tech and Reference details

I am using Genetic Programming written in Java, the program is based on Laurence Ashmore's Master thesis, as well as his supervisor Miller who pioneered Cartesian GP that allows for genetic drift and is well suited for visual work. This work is using Ashmore's Java code, I have only slightly modified it for this series. Unlike most creative GP systems where a user is the fitness function, this systems uses a resemblance algorithm as fitness, so it runs automatically. I am able to pick interesting individuals during the runs and save both their genetic code (so I can reuse it or combine them later - HERE - or render higher res images) and the thumbnails here. These images use a more painterly Hue, Saturation and Value (HSV) color space as well as a function set that works well for images. For additional info on the GP work, check back here (it is still running now) and see Ashmore's site: www.emoware.org and Miller's site: www.cs.bham.ac.uk/~jfm/ For a good overview of GP see http://www.genetic-programming.com

I am grateful to Ashmore for his work and support.

All images are un-retouched and at high compression and low resolution.?