[WhatDoYouThink?]

I have already thought through this problem.

Any problem space the NN is to discover  generalisations e.g. what are the charactiestsixs makes a dog a dog.

The difficult of this is unknown until you run it through an NN ..if N layers is not enough add another layer until you get accuracy you want.

So one needs to figure out the "ease of generalisation" to determine the structure of the network needed. 

I can do this by measuring the error vector of after training an X(say 3) layer NN.
and using the error distribution I trained another network to predict the #of layers needed to achieve a desired error.

that's only the numbers of layers. still got numbers of neurons, their interconnections (mostly used all-to-all), ...... a lot of guess work and not optimised