That's awesome! I've been playing around with some machine learning stuff recently, and because of my art background, I've been wanting to tackle similar stuff to this. Did you let your genetic algorithms run on real-life paintings? i.e. Did they get refine after making a physical model, or did you do the refinement and evolution before it got the physical creation phase?

The GA was based on an open loop simulation. I tuned a digital canvas / painting simulation by hand and then evolved the paintings in simulation. After evolving the painting I would execute the painting with a robotic arm. I did some preliminary work on closing the loop: using the actual output of the painting to tune the simulation but didn't get too far. A paper on my methods / simulation and results: http://carlos.ag/GA2008paper.pdf