That's been tried before. It's theoretically possible, but unless you have a ton of computing power and example problems to work on, it'd probably just overfit to the specific problems you train it on.

And as the other comment pointed out, genetic programming is far too weak to achieve anything this complex. You need to simplify the problem tremendously by hand and then have the computer optimize the last few bits. Though seeding existing metaheuristics to start with might work.