Hi,

My name is Gal and I'm a second degree student in EE.
I'm working on an RBF neural network to solve a function approximation problem.
The big question I have is this:
In order to find the optimal number of hidden neurons, I was thinking of using the cross validation method and choose the optimal number of hidden neurons as the one that gives me the smallest error on the validation set.
But for each network with some fixed number of hidden neurons I get a different validation error, that depends on the initial value of the centers of the RBF.
So, for example, if the number of hidden neurons is 10, I can get different validation errors, depending on the initial values of the centers (the centers are trained according to the K-means).
So, I think I should run say 20 times on each number of hidden neurons and pick the network that gave me the smallest validation error (after choosing that network I will be able to use it and find the generalization error on a test set).

Does it make sense?
I heard another opinion saying I should pick the network with the lowest mean validation error.

What do you think?

I'll appreciate a fast reply.
Thanks!
Gal