I think

It is of use to research the performance of nonuniform shapes.we often use uniform microstrip lines or stripelines to synthesis couplers or power splitters,
but they all have limitations especially in wideband application.
I think we should apply genetic algorithms and artificial neural networks to the study of the shapes and to the synthesis of new components .
Perhaps more astonishing invents will appear

your thinking is right the neural network and genetic algorithms are already used for such application for example see

http://web.doe.carleton.ca/~qjz/qjz.html

Well what you are suggesting is quite usefull concept , but the problem i see is that it is not going to be of any practical significance so soon.
As the two methods you suggested are still in there developing stage , and always remember the rule of maturity
"One immature technology can seldom lead to another matured technology "

But yes as we have seen things have started happening (as in above link eg.) . It holds a promise...but till then problems stay ! Go for traditional options !