(posted here instead of one level up because it seems to expand on the thoughts of the parent)

++ to Limbic~Region for the algorithm (grandparent node). I like it.

It does appear, however, that it is minimizing the error once for each individual fragment, rather than minimizing the error for some summation function (absolute error, statistical deviation, etc) over the entire population.

As a "good" solution, I think that this is excellent, and it seems to be similar to a "good" solution to the bin packing problem that I remember from my undergrad days. It can be fooled with some nasty data.

Like any other NP Hard or optimization problem, an approximation's "good enough" value is in the eye of the beholder, and if you are willing to calculate (or wait for) the optimum solution.

You could add an optimization loop to the end of this algorithm, and bound it with whatever time you are willing to wait (or some other limiting factor), and start swapping values from the bins (picking which ones will give the best change is left as an exercise - perhaps swapping something from the best and the worst, or the two worst, or the worst over and the worst under ), recalculating the fitness, saving the state if better, and repeating if you still have cycles left.