The trick is, the optimal trade-off between variation and selection pressure is highly dependent on the problem domain (selecting to maintain variation is great when navigating a broad fitness landscape with many sparsely distributed peaks of different sizes, whereas strong selection pressure is best when climbing to the top of one of those peaks). I'd ideally like to pick out a way to embed this trade-off inside the evolutionary process itself, so that individual selection could implicitly seek out a good balance between variation and hill-climbing, based on the current complexity of the problem being solved. I think allowing for some sort of speciation (to some extent, speciation boils down to running natural selection on groups of naturally selected individuals - the recursive possibilities are clear, here, though I don't know if they're useful) as well as allowing the genetic material to (in some manner) specify how it "uses up" its fitness is one approach to this, but I haven't settled yet, that approach still seems a bit ad hoc to me and I'd really like to find a more natural way to achieve something similar.

As a rule, I definitely think that playing with the rules of evolution, and more specifically, letting those rules (perhaps implicitly) evolve themselves, will be a key idea if we want to exploit evolution for all it's worth; finding a way to let the process significantly improve itself is, IMO, the Holy Grail of evolutionary computing. Evolution is, in some sense, a process of emergence that's been optimized by emergence, and I suspect that our best shot at creating intelligence may be to follow that pattern and attempt to achieve an evolutionary process that's optimized by evolution.

What results when you have intelligence optimized by intelligence is another matter, and (IMO) a far more dangerous one, but that's another story for another day...