[ADMB Users] a difference in estimates between versions?

[Allan.Hicks at noaa.gov]

I think that we are narrowing in on the issue. I have not been able to
get a positive definite Hessian with MS 64-bit Windows compiler and ADMB
10 (I am not sure if it is 10.0 or 10.1).  However, with the 32-bit
compilation, I got similar results as others (r~0.4999). I am wondering
if the problem is false cnvergence and the 32-bit version is giving a
local minimum.  The .par file from the 64-bit versions shows a better
likelihood but no Hessian.

MS Win 32 (notice max grad)
# Number of parameters = 4  Objective function value = -5.07537  Maximum
gradient component = 540.555
# r:
0.499949443583
# B1:
88781.9987683
# K:
91364.7101759
# lnq:
-2.16787320600


MS Win 64 (-maxfn=10000 but no Hessian)
# Number of parameters = 4  Objective function value = -5.52188  Maximum
gradient component = 0.000491713
# r:
0.558207609664
# B1:
101608.783998
# K:
83033.1712739
# lnq:
-2.08649743192


I have not looked thoroughly at the model, but maybe some tweaking will
help convergence.

Any thoughts?
Allan




----- Original Message -----
From: Richard.Methot at noaa.gov
Date: Wednesday, October 5, 2011 12:11 pm
Subject: Re: [ADMB Users] a difference in estimates between versions?

> Something that runs for 1556 iterations to estimate 4 parameters is 
> probably on a very flat likelihood surface, so the chance of small 
> numerical differences making noticeable differences in the results 
> is 
> not surprising.  However, the ~10% CVs reported in Derek's results 
> suggest more curvature in the surface.  Perplexed...
> 
> Rick
> 
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/users/attachments/20111005/118c09fa/attachment.html>
-------------- next part --------------
_______________________________________________
Users mailing list
Users at admb-project.org
http://lists.admb-project.org/mailman/listinfo/users
-------------- next part --------------
_______________________________________________
Users mailing list
Users at admb-project.org
http://lists.admb-project.org/mailman/listinfo/users