[ADMB Users] Odd difference between 10.1 32 bit safe and optimized mode

[James Bence]

I have been working with  a colleague on an assessment and they brought the
following problem to my attention.  They had been using admb-IDE on a
windows machine and had routinely running versions of the assessment with
no apparent problems (exit code 0, reasonable Hessian and standard errors
etc) and there were no run time errors, but decided to rerun things and get
some additional output but forgot to turn on safe mode.   When they did
this the model failed to converge properly (stopped when failing to improve
likelihood) and got Hessian warning (exit code 1 and no var-cov matrix).

After some exploration we found that this difference in behavior occurs for
the 32 bit version of the release distributed with admb-IDE for windows but
not the 64 bit version of 10.1 distributed with IDE.  The 64 bit version
produces exactly the same crappy results we got for the 32 bit optimized
version, whether it is run in safe mode or optimized.

If the optimized 32 bit mode version is started from where the safe mode
converged to it does not spin into outerspace but also produces a converged
solution, var-cov, exit code zero with parameter estimates etc nearly
identical to where it started.

So with that background my question is whether the different paths to a
solution and results between 32 bit safe and optimized is a known issue?  I
had thought the only difference between safe and optimized had to do with
whether array bounds were checked but it seem clear that for 10.1 32 bit
there is something else different too.  Was this caught and fixed for 11?
Note I am not asking for help getting our assessment to work for all the
admb versions.  It appears that our original starting values were just too
out of whack for the 32 bit optimized or 64 bit versions but for some
reason the 32 bit safe mode version dealt with it (for most model variants
anyway).  We can be more careful with starting values (or try other things
to make the fitting more robust) but what I am really concerned about is
whether there are unintended differences between the safe and optimized 32
bit 10.1 (or other releases) admb.   Is this a reason to stop using 10.1?

Thanks,

Jim Bence
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/users/attachments/20140203/835e083a/attachment.html>