[MMTK] How to check a minimized universe for convergence

Konrad Hinsen hinsen at cnrs-orleans.fr
Mon Jun 29 06:21:54 UTC 2009


On 26.06.2009, at 14:38, Guido Wagner wrote:

> I'd like to check if a minimization did converge and how many  
> minimization steps were used (ConjugateGradientMininimizer ). Is  
> there an easy way to get these values, that does not hurt  
> performance too much ?
>
> I allready thougt of the following methods:
>
> - create a trajectory and read the number of steps from the  
> trajectory-object (might slow down the calculation)

You might try creating a trajectory without storing any variable.  
That would create just the step counter variable, thus minimizing I/O  
overhead. Unless your system is very small, the overhead should then  
be negligible. You could even use a circular trajectory of length  
one, to minimize file size, and read out the step counter at the end.

> - add a counter as a forcefield term (the forcefield is evaluated  
> far mor often than the max. minimization steps)

For conjugate gradient, there are many force evaluations per  
minimization step, so you would be measuring something quite  
different. A CG step is in fact a 1D minimization, with the direction  
changing at every step.

> - create a custom action-object for counting (as far as I found  
> out, this requires writing a special C-library with a Cfuntion just  
> for counting)

Yes.

Konrad.
--
---------------------------------------------------------------------
Konrad Hinsen
Centre de Biophysique Moléculaire, CNRS Orléans
Synchrotron Soleil - Division Expériences
Saint Aubin - BP 48
91192 Gif sur Yvette Cedex, France
Tel. +33-1 69 35 97 15
E-Mail: hinsen at cnrs-orleans.fr
---------------------------------------------------------------------






More information about the mmtk mailing list