Page 1 of 1

Restricting Parameter Domains

Posted: Tue Jun 17, 2014 6:09 am
by shailesh
Is it possible to restrict the domain of a parameter such that it:
> takes only integral values
> takes only odd numbers (subset of above case)

I do remember reading that:
The PRAXIS optimizer used by NEURON requires a continuous objective function, so restricting parameter space can actually interfere with it
So not sure if it just not allowed or not as efficient if we restrict the parameter space. If allowed, how do we go about achieving it?

Re: Restricting Parameter Domains

Posted: Tue Jun 17, 2014 5:01 pm
by ted
PRAXIS is a gradient descent optimizer. It won't work properly if it can't estimate the gradient of the error surface. In other words, PRAXIS requires that the partial derivatives of error with regard to each parameter must all exist and be continuous, which in turn requires that the parameters themselves must be continuous over their individual domains. If there is only one parameter that takes on discrete values, a workaround might be for you to identify a small set of "likely values" which that parameter might assume, then run a separate PRAXIS optimization for each likely value of that parameter. Or you might scour the optimization literature for something that is better suited to your needs.

Re: Restricting Parameter Domains

Posted: Tue Jun 17, 2014 11:34 pm
by shailesh
Thanks, the workaround you suggested would be good enough for me in this case.

p.s. could I bring your attention to another post that I have not yet been able to resolve: