[Rd] convergence problem with nlme() (PR#2369)

Douglas Bates bates@stat.wisc.edu
Fri Dec 13 00:15:05 2002

It turns out that you learn more if you set verbose = TRUE in the call
to nlme rather than setting control = nlmeControl(msVerbose = TRUE).

The verbose = TRUE option shows you that the parameter values are
bouncing back and forth between two regions of the parameter space,
neither of which are close to the optimum.  It happens that nlm will
occasionally take very large steps at the beginning of the
optimization resulting in unusual parameter values.

This example is a difficult optimization problem.  It may be possible
to stabilize it somewhat by replacing the internal calls to nlm by
calls to optim but that brings its own set of difficulties.