Multiple Run Fitter options

Using the Multiple Run Fitter, praxis, etc..
Post Reply
Ginetto
Posts: 14
Joined: Tue Mar 09, 2010 4:50 am

Multiple Run Fitter options

Post by Ginetto » Wed Oct 06, 2010 9:45 am

Dear all,

I have read the tutorial “Using NEURON’s Optimization Tools” but I still don’t have clear some options such as “append the path to the savepoath.fit”, “|| Optimize” and “# quad forms before return”.
Is there another tutorial I can read?

How can I fix the numbers of runs before stopping the optimization?

Thanks for your help.

Best wishes,
Gino

ted
Site Admin
Posts: 5621
Joined: Wed May 18, 2005 4:50 pm
Location: Yale University School of Medicine
Contact:

Re: Multiple Run Fitter options

Post by ted » Thu Oct 07, 2010 1:34 pm

Ginetto wrote:I still don’t have clear some options such as “append the path to the savepath.fit”, “|| Optimize” and “# quad forms before return”.
Is there another tutorial I can read?
Nope. There is only the documentation in the Programmer's Reference and the source code. First try the Programmer's Reference--especially the alphabetical index. If you have gleaned what you can from the Programmer's Reference, find your way to NEURON's hoc library (see Secrets of NEURON: the hoc library viewforum.php?f=28), then start grepping for strings of interest and reading the files that contain those strings.**
How can I fix the numbers of runs before stopping the optimization?
You can't. That's not how the MRF works. The MRF's GUI allows you to specify the number of quadratic forms that the praxis optimizer finds* before returning, or whether you want to let the optimizer decide for itself when to return. hoc has a function called attr_praxis() that allows you to specify two parameters called tolerance and maxstepsize; see the Programmer's Reference for more information. In lib/hoc/mulfit there is a file called fitparm.hoc that uses attr_praxis() to set tolerance to 1e-4 and maxstepsize to 0.5.

*--The praxis optimizer uses Brent's principal axis method to approximate the local shape of the error surface by a quadratic form ("a homogeneous polynomial of degree two in a number of variables"). A quadratic form has a point at which its value is minimum. The hope is that, by doing this one or more times, the sequence of minima will converge on a minumum of the error surface.

**--If you do, here's one thing you'll discover: savepath.fit is a file to which the MRF will write a record of parameter values and corresponding values of the objective function as it hops from quadratic form to quadratic form.

Post Reply