I just noticed that MAPS 3.0.4 finds (slightly) different truncations on the very same structure set when started with different -g parameters.
Is this intentional behavior? And if so, could someone explain why that is?
I don’t understand what the reasoning behind this would be, as it is my understanding that the supplied -g parameter just determines the initial set of predicted structures (as referenced in predstr.out) and therefore should not interfere with the actual fitting procedures.
To reproduce this finding, all that needs to be done is to call MAPS with different -g parameters, e.g. calling it via
maps -sig=10 -c0=0.0 -c1=0.5 -g=80 -2d
and
maps -sig=10 -c0=0.0 -c1=0.5 -g=64 -2d
on the structure set attached to this post. In this case, MAPS finds the same choice of clusters but different ECIs, see the eci-g{80,64}.out files in the attached tar.gz-file.
Regards,
stichr
EDIT:
The file upload seems to be broken, so I uploaded my data here.
Your are right that the CE should be independent of how many structures there are in maps’ internal database of structure whose ab initio energy is not known.
What is happening here is that your system has a lot of near ground states very close in energy and maps has to weights some of these structures more heavily in the fit to make sure their ordering in energy is correct. The algorithm to assign these weight is not completely independent of the order in which the structure are stored in memory (it tries the first problem it finds, fixes it, etc. until no problems are found - sometimes trying the fixes in different order needs fewer tries to fix all problems).
Here, when you generate a larger internal database (-g=80) all your structures are generated in maps logical order and then it looks at your manually entered structures and puts them it that order. With the smaller database (-g=64) some of your manual structures are not generated by maps first and are just put at the end of the list.
You can see that is what is happening because the weights (column 5 in fit.out) are different in both runs. Also, the ordering of the structure is a bit different.
Note that differences between the two CE are fairly small, so I would just ignore the "problem", which is not really a problem.