Following Daniel's suggestion I added a set of Movers to my MC optimizer (see code attached). Now the optimizer actually does something, though the results are very strange. It is much probable I poorly chose my parameters, and 1000 steps are not so much for a MC optimization.
It also appears I have a problem with the cost function : it evaluates to 18.36 for the native configuration and to 1371 for an initial "noisy" configuration not so far from the initial configuration. It then evaluates to 1.03 for an "optimized" configuration where all four chains are placed far outside the sampled density. I did not discover yet if and how it is possible to set parameters for the global cost function, maybe my problem has something to do with it.
Anyhow, I would be very grateful if someone would glance at my script, just to check I did use IMP in a proper manner.
--Ben
Le 23 juil. 2010 à 19:44, Daniel Russel a écrit :
> The bug sounds like something Elina found and patched after the 1.0 release. I'll test your script with the current SVN. > > As general comments: > - with monte-carlo, you have to provide the move set. We really should put up an error if there is no move set :-) I'll fix that. > - Conjugate gradients can get stuck very easily (which is why we are working on a better optimizer). This is especially true with rigid bodies composed of my particles as the atom balls can get mixed in together and CG can never separate them.
Biocomputing group Email : schwarz@igbmc.fr Voice : +33 (0)3 68 85 47 30 FAX : +33 (0)3 68 85 47 18
Structural Biology & Genomics Dept. - IGBMC 1 rue Laurent Fries BP 10142 F - 67404 Illkirch CEDEX
