[RMCProfile-users] Weighting of data sets

[Henrik Mauroy]

Dear Graham.
Thank you so much for the clarification of the weighting(=sigma)!
I didn't know RMCProfile worked without Bragg-data, but I now saw that 
it works like a charm when testing without the Bragg section.

Cheers,
Henrik

On 03.06.2014 21:01, King, Graham M wrote:
> The simplest solution to your problem would be to simply not include the Bragg data at all and only fit to G(r) and F(Q).  RMCProfile does not require you to fit Bragg data, you can choose whatever data types you want to fit.  In fact, you can run it using no data at all if you wanted to do something like randomly swap atoms or disorder them.  Just take the entire Bragg section out of your .dat file.
>
> A higher value for weight allows more moves that make the fit worse to be accepted.  The weights are supposed to be related to the actual errors in your data, to help prevent you from over-fitting the data.  So you could also make the weight value high for your Bragg data and then no moves would be rejected because of it.  From my experience weight numbers between 0.02-0.06 seem to work best, but it also seems to depend on the exact situation.  The weights are the sigma values used to calculate the total X^2 for the fit, as described in appendix C of the manual.  You should read this section for a better understanding how the weights you set affect the refinement procedure.  The section describes the mathematics of it.  The X^2 is the number the RMC is trying to reduce, and it is inversely proportional to sigma^2.  Since the probably of accepting a move that makes the fit worse is lower for larger X^2, using lower weights allows more atom moves to be accepted.
>
> -Graham King
>
> ________________________________________
> From: Henrik Mauroy [hmauroy at gmail.com]
> Sent: Tuesday, June 03, 2014 12:24 PM
> To: rmcprofile-users at rmcprofile.org
> Subject: [RMCProfile-users] Weighting of data sets
>
> Hi everyone.
> This might be a dumb question but I wonder how the weighting of the data
> sets are done. So far I've used equal weighting for G(r), F(Q) and Bragg
> and haven't been troubled by this yet.
> I have some data collected at GEM where the Bragg data is completely
> impossible to fit due to a unusually weak Bragg signal (due to
> cancelling negative scattering lengths), and need to turn the weighting
> of the Bragg data to such a low number that it is not being part of the
> fitting.
>
> 1) What effect does a higher number for the weight do?
> 2) In what range should the weight-number be? In most examples it is
> smaller than 0.1, but can you use bigger or smaller numbers?
> 3) I see from experimenting with high and low numbers that the chi^2
> value calculated for the data set in question becomes larger for a lower
> weighting. Is this the number RMCProfile is trying to reduce?
>
> Thanks for any answer!
>
> Cheers,
> Henrik
> _______________________________________________
> rmcprofile-users mailing list
> rmcprofile-users at rmcprofile.org
> http://lists.rmcprofile.org/mailman/listinfo/rmcprofile-users
> --
> Scanned by iCritical.
>
> _______________________________________________
> rmcprofile-users mailing list
> rmcprofile-users at rmcprofile.org
> http://lists.rmcprofile.org/mailman/listinfo/rmcprofile-users