[eu-gene] Aesthetic Evaluation Algorithms And Fitness Functions

[Rob Myers]

Following on from recent debates...

The algorithmic evaluation of art has a long history. I'm partial to
Gips & Stiny or to Birkhoff myself. I have a toy system inspired by work
I did on the former, featuring a bot that posts simple image
descriptions and a bot that posts simple evaluations of them. The bots
really don't need us to read them, as there is a third bot that watches
for the most highly evaluated works:

http://robmyers.org/art/cybernetic_artworld/

But if I made them more complex, I expect that they would quickly
produce things that made sense to them but not to me. I expect the
problem with using aesthetic evaluation algorithms as fitness functions
for genetic algorithms to be that they will either generate inhuman
aesthetics or sub-optimal human aesthetics.

Aesthetics is a matter of human perception. The measure of art is
whether it looks good *to a human being*. If we create inhuman systems
to evaluate and manage or exploit aesthetics then we cannot be sure that
they will share human aesthetic values. They may be as high frequency
trading is to actual economic activity, or Skynet would be to actual
nuclear security.

Assuming that aesthetics is a complex matter of the human mind, nothing
less than AGI will allow fully competent algorithmic aesthetic
evaluation. Systems that don't know about art but know what they like
are allegorically and philosophically interesting and may be art in
themselves, but will produce ugly antennae, become experts at spotting
cloudy days rather than Soviet tanks, or tune into the noise from nearby
capacitors.

I would love to be proven wrong in this. Not least by my own future
work. ;-)

- Rob.