How do you know how many units to include in the non linear estimator for NLARX modelling ?

Hi all,

It seems as though choosing the number of units for the non-linear function in the non-linear estimator

(sigmoid in my case) for NLARX estimation modelling is a trial and error process. I was wondering if there

is any function or logical way of knowing how many units will be best to include in the non-linear function ?

And what does it mean really ? is it the more the better refinement or the less the faster ? i dont know. Any input will be greatly appreciated

Kind Regards, Steve