For those issues with optimization methods (optim, optimx, and others) I see, a
good
percentage are because the objective function (or gradient if user-supplied) is
mis-coded.
However, an almost equal number are due to functions getting into overflow or
underflow
territory and yielding quantities that the optimization tools cannot handle (NA
or Inf etc.)
Two general approaches I find helpful:
1) even if there are no actual bounds on parameters, put in
"reasonable" limits. They
don't need to be too tight, just enough to keep the parameters from giving a
silly
objective function
2) do some evaluations of the objective to make sure it is really being properly
calculated. Never hurts to have some "known" outcomes.
Beyond this, we get into reparametrizations. Great idea, but far too much work
for most of
us, even if we work in the field.
Best,
JN
On 01/17/2011 06:00 AM, r-help-request at r-project.org
wrote:> From: Uwe Ligges <ligges at statistik.tu-dortmund.de>
> To: Jinrui Xu <jinruixu at umich.edu>
> Cc: r-help at r-project.org
> Subject: Re: [R] fgev_error_matrix_singular
