[ADMB Users] -lprof

[John Sibert]

Ugh.

The variable I'm trying to profile has a pretty low estimated standard 
deviation. From the .std file:
>     87   natural_mortality 1.5239e-01 7.8301e-03
By fiddling with the model structure and the lprof control variables, it 
will sometimes work, but the resulting profile plot is pretty strange. 
See attached.

I'm working to track down the problem, but even simplifying the model it 
is slow going because of the need to compute the Hessian every time.

The contents of diags is:
> f$ cat diags
> tempint1
>        inf       inf       inf       inf       inf inf       inf       
> inf       inf
> m(2)
>        inf       inf       inf       inf       inf inf       inf       
> inf       inf
> m(3)
>      0.882     0.956     0.989     0.999         1     0.999 0.989     
> 0.956     0.882
> xdistance
>  1 1 1 1 1 1 1 1 1

Changing exp to mfexp for the computation of tempint1 in newmodmn.cpp 
seem to help a bit, but does not solve the problem. I think the cause 
occurs earlier, but I'm having trouble tracking it down.

Does anyone really use this option? It seems like a good option to 
diagnose an ill-defined parameter when MCMC would take too long.

Would it discomfort other users to change the way the procedure composes 
the file name from the variable name?

John

John Sibert
Emeritus Researcher, SOEST
University of Hawaii at Manoa
Honolulu HI (GMT-10)
808-294-3842 (mobile)

Visit the ADMB projecthttp://admb-project.org/

On 07/17/2014 06:03 PM, dave fournier wrote:
>
> Almost anything might go wrong with lprof.  It works by maximizing the 
> log-likelihood subject to a constraint
> on the value of the variable being profiled.  This process drags the 
> values into what might be areas where the
> function becomes unstable in some way.
>
> Hard to say more about a particular case without seeing it
>
>      Dave
>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: allskj-p001-natural_mortality.png
Type: image/png
Size: 32110 bytes
Desc: not available
URL: <http://lists.admb-project.org/pipermail/users/attachments/20140720/eb7dab54/attachment-0001.png>