[eu-gene] need advice on a framework design
list at philipgalanter.com
Tue Jan 24 22:47:08 GMT 2012
On Jan 24, 2012, at 3:35 PM, David Hart wrote:
> My gut reaction is that even if an evaluation system could be developed that allowed human evaluation to be taken out of the evolutionary loop, it would probably turn out to be very specific to the peculiar evolving expressions look. So the result wouldn't generalize and wouldn't really yield much leverage on or understanding of evolutionary art and design of other kinds. But again, that's just an intuition and not something I can demonstrably prove.
> The theory claims otherwise, so if its impossible to escape a narrow range, people using the technique should be asking why. I acknowledge that its *difficult* to escape the look, but I disagree that its impossible, and I think I'm providing some small amounts of evidence to back up my claim with some of my results, for example:
My point here isn't about the limited (or not) aesthetic range of evolving expressions. It's about whether a corresponding automated (non-human non-interactive) fitness function will have any value beyond its very specific application to evolving expressions. Assuming, of course, that such a fitness function can even be created. A number of attempts have had limited success.
> I'm highlighting these particular samples for a very specific reason, not because I like them. They are examples of escaping the usual obvious tells, and bunch of them demonstrate some mild success of the evolutionary process by appearing to have shading and three dimensionality. I selected for those attributes, it is not some inherent part of the building blocks I used. I'm using the same building blocks everyone else is, and sometimes selecting for shape and shading, and sometimes achieving it. This suggests the evolutionary theory is working, as well as suggesting that it is in fact possible to escape the well known look, even if I didn't get all that far.
Well the proof of the pudding will be in the tasting and so as soon as evolving expressions yield a wide variety of personalized graphical styles I'll be happy to concede the obvious.
Frankly I don't think "quantum leap" improvements in evolutionary art, whether using expressions or some other kind of genotype format, can be made without solving one or both of the following problems. There are lots of caveats and equivocations that can be applied to these problems, but I'll state them simply here:
(1) Industrial applications of evolutionary computing are typically single or multi-dimensional optimization problems where an unambiguous and tractable fitness measure can be calculated for every phenotype. Because computational aesthetic evaluation remains an unsolved problem, and fitness scores have to be supplied manually by humans, evolutionary art does not have the same advantages of large populations and generations for development.
(2) The direct mapping of genotype into phenotype in a single step severely limits the complexification capacity of evolutionary art systems. The complexity of life in nature is due, in part, to the fact that genes are expressed through many levels of emergence. I.e. DNA to proteins to organelles to cells to tissues to organs. This is why current evolutionary art can only produce the equivalent of variations of fish, but cannot to date produce the equivalent of fish evolving into amphibians into reptiles into mammals.
(I don't mean to be presumptuous but those who would like to read more about this are welcome to check out my paper: The Problem with Evolutionary Art Is . . . ).
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the eu-gene