[ADMB Users] Please test updated installation documentation - Bug still not fixed

[Alex Campbell]

On Fri, Nov 22, 2013 at 1:01 PM, dave fournier <davef at otter-rsch.com> wrote:

>  On 13-11-21 06:29 PM, Alex Campbell wrote:
>
> I'm not sure I understand this.
>
> what does this mean?
>
>
> mceval output inconsistent with mcmc display"
>
>
>
It is the subject line of the email to this list in which I first reported
this issue.



> I know I wrote the mcmc stuff really fast and wanted some output that I
> could plot with gnuplot
> for a useless  meeting in 1992.
> It worked fine with a good model, but not with a "bad" model.  The mceval
> stuff was really meant
> to supercede it i.e if you got good results with the display fine, but the
> real output was intended to
> be the mceval stuff.
>
>
Actually the mceval stuff will be wrong too if you have an unidentifiable
model on windows/mingw, complete with a bogus std file. On a real system,
ie linux-gcc, once the estimation phase is complete, it will give a
diagnostic message and quit. No std file will be produced.


> I wonder what happens with R if you write a model with unidentifiable
> parameters?
>
>
Good question.


> There is a standard thing I use (or used to use back when I was still
> alive) I would put a small quadratic penalty
> on all the parameters when the problem seemed degenerate. with a little
> bit of tuning it was usually
> possible to find where the unidentifiability was.
>
> Anyway if you are trying to convince a someone of the value of ADMB, I can
> think of better ways than
> trying to fit a model with unidentifiable parameters.
>
>
>
In theory you never build models with unidentifiable parameters. In
practice, at least in my experience, it comes up regularly, especially if
you're playing with spatial stuff. This is a problem if you can't rely on
your tools to tell you you're being a blockhead. You say "when the problem
seemed degenerate".. the issue is how do you know this from ADMB output? If
you have a degenerate model-data combination, on a windows system with
mingw all bets are off. If you aren't experienced enough to guess that
degeneracy / unidentifiability may be the issue ADMB won't tell you. There
are other, subtler discrepancies between ADMB on windows and linux. At work
I'm forced to use windows but for stock assessments I will always cross
check with a linux box. I know you guys are putting a lot of work into
making things easy for new windows users, which is great.. I think ADMB is
an amazing piece of kit. It is just unfortunate it is being (or at least
has been in the past) let down by a substandard operating system and/or
compiler.

You could include this in the testing. Just add a misspecified model-data
combination to the examples.


>
>  I posted on this a couple of years ago. See "mceval output inconsistent
> with mcmc display". The last word there is me saying I've figured it out
> and I'll send you the code to reproduce the issue off list. Can't remember
> if I did, and don't have access to those emails to check. I think I did
> send it to you and never heard back. I assumed since no one else had come
> across this since that time it might have been resolved with updates to the
> mingw compiler. But no. Came across it again about six months ago on
> another model, ADMB version 10 i think or maybe 10.1. I was in the process
> of winning over a colleague to the ADMB cause and had to basically say well
> its great but its broken on windows. We ran the same code, same inputs on
> linux and it diagnosed the unidentifiability (can't remember the exact
> message but i assume it was telling us that the hessian was singular to
> machine precision). We had an unidentifiable virgin recruitment in one
> spatial region, which made sense in hindsight. If you can't find the code
> and inputs I sent you in June-July 2011 (or if I didn't?) I can dig around
> for the more recent model. Can't remember if I've checked this out on a
> visual studio compiler but I have a hunch the issue is mingw not windows in
> general.
>
>
> On Fri, Nov 22, 2013 at 1:19 AM, dave fournier <davef at otter-rsch.com>wrote:
>
>> My experience of the mingw compiler is that with it ADMB fails to
>> diagnose singular or near singular hessians. This is a minefield for
>> inexperienced users and in my opinion it should not be supported. Has no
>> one else seen this issue?
>> Alex
>>
>>  That seems highly unlikely in the sense that the mingw compiler would be
>> different
>> from any other compiler in this manner. Do you have an example?
>>
>> _______________________________________________
>> Users mailing list
>> Users at admb-project.org
>> http://lists.admb-project.org/mailman/listinfo/users
>>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/users/attachments/20131122/19f66dea/attachment.html>