The file omxSadmvnWrapper.cpp has the following hard-coded precision options for numerical integration:
double absEps = 1e-3; double relEps = 0; int MaxPts = 10000*cov->rows
The last of these deviates from classic Mx setting of
1000*ordCov->rows * ordCov->rows * ordCov->rows;
We have been experiencing persistent difficulties with optimization when the number of ordinal variables is large - more frequently than with classic Mx. Possibly the MaxPts setting could be changed to agree with classic Mx?
At the same time, providing mxOption() access to absEps and relEps would help to eliminate possible numerical precision related issues. Could that be added?
Comments
#1
Joshua Pritikin added this feature but it has not yet been tested or documented. I can certainly try to test it (possibly with help from Michael Neale), once it is documented. Joshua, can to update trunk/R/mxOption.Rd for this feature?
#2
See svn 3213
#3
I'm closing this. I believe it's implemented and documented. If someone feels otherwise, they can re-open.