Wavelets are my favorite approximate function to build up with.

But yeah, the overfitting problem / approximation is the real issue. I do like seeing the hybrid approaches (Genetic Algorithsm searching random Neural Nets with a little bit of backpropigation for some measure)