Programming Praxis' stop criteria

Using the Multiple Run Fitter, praxis, etc..
ted
Site Admin
Posts: 5624
Joined: Wed May 18, 2005 4:50 pm
Location: Yale University School of Medicine
Contact:

Re: Programming Praxis' stop criteria

Post by ted » Mon Nov 24, 2014 12:27 pm

Weaver and Wearne have written code that adds simulated annealing to the MRF--see
https://senselab.med.yale.edu/modeldb/S ... odel=87473

However, why don't you first try some optimizations that start from different initial points in parameter space (e.g. corners of the "hypercube" that bounds parameter space)? If they all end up near each other, that would reduce concern about the possibility of local minima in the error surface.

mart
Posts: 13
Joined: Tue May 06, 2014 11:59 am
Location: Physiology Department - University of Bern

Re: Programming Praxis' stop criteria

Post by mart » Fri Dec 05, 2014 9:51 am

I tried to start my Optimization process with MRF from different points in the parameter space.
In many cases cases (6 out of 10) the optimization finished with a very similar trade-off, a set of values for my free parameters that minimized the error to something like 11...
These set of values were fairly similar among all the 6 optimizations. In the rest of the cases the simulation got stuck in a local minimum with a way larger minimum error, something like 23.

Thanks a lot for recommending Weaver-Wearne's strategy, It's wonderful!

I have tried only once, but it looks great. It kept my computer running simulations for 3 days while exploring the parameter space in a temperature dependent manner (as expected for a simulated annealing approach) without any manual intervention. When the optimization was finished I got a new minimum error, which is the best fitted version of my model so far, my current minimum error is 9.5. The best set of values for my free parameters was quite similar to those set of values found by my previous optimizations with the default version of MRF, only in one case (out of a total of 4 free parameters) the value was remarkably divergent.

From my point of view and with my basic understanding of simulated annealing, my last results corroborate the hypothesis that MRF was actually performing quite well, it basically pushed my model to the best of its capabilities. Nevertheless, now, I feel much happier with the exploration of the parameter space.

I am starting to seriously consider the hypothesis that the error between model and experimental data comes from voltage offset issues during some double patch clamp experiments... I will repeat these experiments paying special attention when offsetting the electrodes, then I will tell you how everything went.

I was completely ignorant about the existence of the aforementioned Weaver-Wearne implementation of simulated annealing in NEURON, sorry about that. It is a very powerful tool. I don't know how feasible is this, but if there are no major impediments or competing interests, I would suggest simulated annealing to be available by default in the next stable release of NEURON. It could be very useful for lots of users. I would definitely try to help as much as possible in such a project.

Best wishes,
Ulisses

Post Reply