So, I know that facial recognition is pretty decent, but I think that HyperNEAT would be pretty good at doing this. I think that it would be able to take advantage of the geometry of the face in order to better characterize the face. I have found a paper that used HyperNEAT on digits, and it did not do extremely well. They found that HyperNEAT is good at extracting useful features that could then be used to classify. So, maybe that would be something else to look at.
To determine the number of genomes within the population automatically could use F-statistics (http://en.wikipedia.org/wiki/F-statistics). It doesn’t seem like it has been explored in genetic algorithms, rather for human population studies. Another way that subpopulations could be culled is to use the inbreeding coefficient (http://en.wikipedia.org/wiki/Coefficient_of_relationship).