Umh.. twenty?

You'd be applying a weak optimization process to the problem instead of using your built-in much stronger one, and hoping that its different set of biases will let it hit on a useful algorithm that you yourself wouldn't.

Intuitively, math-space is too big and twisted for evolution to work, and it'd suffer horribly from getting stuck on local maxima. I don't know this for certain, however, and even if you fail you'll still have learned something.