Hi All,
I have placed new Windows, MacOSX (for Intel) & Fedora 5 versions of
GSAS on the ccp14 web site. changes are:
Nov. 12, 2007
1. Oops, forgot some arrays - GSAS now WILL do up to 200,000 profile
points in a single powder pattern.
2. A new weight modification for powder data is available. In histogram
editing (same as where histograms are inserted) the "F" option is used
to set the weight factors for the powder histograms. The first value
("weight factor" - FACTR) is a constant multiplier (default = 1.0) for
all the powder profile weights. The second value ("weight balance" -
WTM) changes the relative weighting between the low and high ends of the
pattern (default WTM = 0.0). If WTM > 0.0 then
w' = w*FACTR*exp(-WTM*T) (this will down weight the high T part
of the pattern)
If WTM < 0.0 then
w' = w*FACTR*[1-exp(-WTM*T)] (this will down weight the low T
part of the pattern)
where w is the weight obtained typically from 1/esd**2, T is either
2-theta in centidegrees or neutron TOF in microsec. The weight
modification is applied when POWPREF is run. Non-zero values of WTM can
be used to "force" a refinement to fit one end or the other of a pattern
(at the expense of the other end) to facilitate jumping out of local
false minima. Very small values of WTM are quite effective; in an
initial trial with x-ray powder data, I set WTM=+/-0.001 and then back
to WTM=0 for a sequence of protein refinements.
3. After considerable thought and a read of the paper "A Study of
Cross-Validation and Bootstrap for Accuracy Estimation and Model
Selection" by R. Kohavi, Intl. Joint Conf. on Artificial Intelligence
(1995), I have reinstalled the facility for producing an "Rpfree" for
powder data. It was originally introduced on Apr. 15, 2004 (it was
removed July 25, 2005). The note given at the time of introduction was:
2. There is now an option in histogram editing to set aside a certain
fraction of the profile points in a powder pattern to calculate an Rfree
and Rwfree. This replaces the old 'S' option (which didn't work anyway).
The Rfree & Rwfree values appear in the .LST file (not on screen) and in
the EXP file. In my initial tests these seem to follow the Rp & Rwp
values closely through the entire course of a refinement but maybe you
all will see wilder behavior in your refinements. Note that you set two
things for this - one is the percentage of points to set aside (1-10%)
and a random number seed. The seed ensures that the selection of points
remains the same for all runs of POWPREF; this will change if seed is
changed or if profile limits & excluded regions are changed. If you want
to use this it should be set very early in the analysis; setting it for
just the final refinement cycles somewhat defeats its purpose.
The maximum fraction that can be now set aside is 50%; powder data is
usually very overdetermined in comparing number of profile points to
number of parameters so this is not unreasonable. I have further
modified HSTDMP and POWPLOT so that the Rpfree points are easily seen
(plots as blue crosses while those used in the refinement are red). It
should be noted that this "holdout" method is not strictly speaking a
"cross-validation" according to Kohavi (but then again neither is the
"Rfree" commonly evaluated for protein structure refinements).
I hope you find these useful. Please let me know if you use these & what
your experience with them is. As usual pleasle let me know about bugs &
"features".
Best,
Bob Von Dreele
