While you play around with the how the algorithm parameters affect its ability to find the polynomial coefficients, it would be nice to see how adding some random noise to the sample points affects it, as well. This would make your example more realistic (because it would mean you'll never be able to generate a perfect fit).
Having used (and still) genetics algorithms to build a good approximation function for a natural phenomenon, I found that I needed to spend a lot of time crafting my chromosomes structure so that it would get a better chance of eventually solving the problem. I also had to put correlation in the fitness function because it would otherwise get stuck in a local maxima where it completely eliminated most of the inputs but the resulting curve looked nothing like the targeted one.
However I am still not completely satisfied with the approximations I get and I would be VERY interested in recommendation for more advanced reading on the subject. Anyone?
