[ADMB Users] non-convergence and "inner maxg = 0"

[H. Skaug]

Hi Ian,

The gradient jumps from 5.48908e+28 to 0 (exactly). That should happen
more gradually.
I suspect something is wrong with your model. See suggestions in
"3.8 Building a random effects model that works" in the Re-manual.
Also, I would start with someting really simple.

Hans

On Fri, Mar 19, 2010 at 5:00 AM, Ian Fiske <ianfiske at gmail.com> wrote:
> Dear ADMBers,
> Thank you for freely providing such an amazing piece of software.  I'm super
> excited about finally having access to all of ADMB's power and flexibility
> and hope to eventually incorporate it into my R package, unmarked, to
> replace R's optim as the engine.  Now, my question...
> First, I fit a fairly complex model in ADMB and it works great.  Next, I
> tried to add random effects and use ADMB-RE, but ADMB-RE appears to not
> converge.  The oddity is that when the program prints out (via cout
> statements) the parameters, they appear to stabilize to the same values as
> in the non-RE version.  And (omitting the parms), the program prints
> ... [I think phase 2 starts here] ...
>   Inner second time = 5.48908e+30  Inner f = -3.14161e+61
> Newton raphson 1
>  inner maxg = 5.48908e+28
> 0
>  inner maxg = 0
> 0
>  inner maxg = 0
> 0
>  inner maxg = 0
> 0
>  inner maxg = 0
> which continues seemingly indefinitely (at least several minutes), printing
> "inner maxg = 0".  If the inner maximum gradient is 0, then why isn't
> another outer step taken or convergence declared?  Instead, it looks as if
> the inner optimization just repeats indefinitely.  That is, I never get to
> Newton raphson 2.
> I appreciate any ideas as to what is going on here or why the optimizer is
> stuck with maxg = 0.  I'm using ADMB 9.1 on Ubuntu 9.10, 64bit.  Let me know
> if you could use any more details regarding the model or my system.
> Thanks!
> Ian
>
>
> --
> Ian Fiske
> PhD Candidate
> Department of Statistics
> North Carolina State University
>
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users
>
>