Restraints, Penalties, and Simulated Annealing

btrump1

I've been running various forms of simulated annealing, using a combination of bond/angle restraints, rigid bodies, and AI_Anti_Bump penalties. More specifically I've noticed that the penalties are much smaller when it first finds the best statistics than when the simulation ends. It's not clear if this is due to the number of penalties I have or some other process is at play during simulated annealing, but I've seen hints of it occurring through a few different *.inp files.

This may also be related to: http://topas.dur.ac.uk/unb/forum.php?req=thread&id=680 . I'm not sure if anyone else has come across this as well.

Thanks in advance!
-Ben

magnuss

Sorry to see that your question has remained unanswered for months. The activity on the form seems to have dropped lately.

It's not clear to me what you mean by "the penalties are much smaller when it first finds the best statistics than when the simulation ends.", but I guess it means that the penalties in your final model are high compared to what you have early in the Simulated Annealing run.

Try putting the keyword
save_best_chi2
in your inp-file.

Without this, the SA will save the solution with the lowest Rwp (i.e. with the best fit to the data with no regards to restraints/penalties), but with the keyword included it will save the model with the best chi2 which has contributions both from the fit and the penalties. If the penalties are still violated more than you find reasonable, you should try weighting them harder.

johnsoevans

My apolgies also for the lack of reply. I try to scan for non-response threads periodically but missed this one. Changes in our local IT service's security rules have meant that the auto-email of new posts to the forum isn't currently possible. This may be why responses are a bit slow at the moment. We have more major changes about to occur. I have my fingers crossed that (a) the wiki will survive and (b) we'll get the email alert facility back afterwards.

That said, I don't have an answer to your question! I think I recognise the symptoms you describe, and may have discussed with Alan at some point but I don't think I ever reached a firm answer (or needed to). I guess it's to do with how relative weighting of penalties vs data evolves. There's a bit on this at http://topas.dur.ac.uk/topaswiki/doku.php?id=penalties_and_restraints&s[]=penalty&s[]=weighting.

My only suggestion is to contact Alan directly if there's a major issue.

alancoelho

The value of a particular penalty at the minimum of Chi2 is dependent on the curvature of the penalty, where, curvature simply means the value of the second derivative of the penalty with respect to independent parameters. Equations 4-4 to 4-7 in the technical reference explains see http://www.topas-academic.net/Technical_Reference.pdf; in particular pay attention to KP.

This means that the value of penalties as a function of Chi2 changes as Chi2 changes. The penalties may well increase early in a refinement and decrease later on, or visa versa.