[ADMB Users] Reading in random effects start values from .pin file

[Ben Stevenson]

Hi Mollie,

Good spotting -- this explains why things are different when a separable
function is used.

My main concern was that ADMB is not doing what is advertised in the
ADMB-RE manual, as you have pointed out. As far as I am aware, it's
supposed to read the .pin file regardless of -noinit, and regardless of
whether or not you have a separable function. I stumbled upon the strange
-noinit behaviour by trial and error while I was trying to work out how on
to get it to read the .pin file.

Thanks for having a look into this; I wasn't brave enough to get into the
source!

Cheers,

Ben



On 6 November 2013 15:11, Mollie Brooks <mollieebrooks at gmail.com> wrote:

> Hi Ben,
>
> I played around with the files a bit and might have an idea.
> If you comment out the exit(100); on line 14 and run it as
> /test -maxfn 1
> and
> /test -noinit -maxfn 1
>
> then you get more output and it looks like the behaviour is similar to
> what you see without a SEPARABLE_FUNCTION (i.e. noinit makes u start at pin
> file values)
>
> It looks like the difference is that when using the SEPARABLE_FUNCTION,
> the first run through the procedure section (and first output) happens
> before the initialization.
>
> Does this seem plausible, or was non-initialization affecting some model
> more seriously?
>
>
> Aside from the difference in behaviour with and
> without SEPARABLE_FUNCTIONs, I would expect the values for u from the pin
> file to be used even when not using -noinit. On page 25, section 3-8 of
> admbre-10.0-rev1.pdf it says
>
> "We can now return to the role of the initial values specified for the
> random effects in the .pin file. Each time step 1 above is performed,
> these values are used—unless you use the command line option -noinit, in
> which case the previous optimum is used as the starting value.
>  The steps it refers to are
> 1. The “penalized likelihood” step: maximizing the likelihood with respect
> to the random effects, while holding the value of the hyper-parameters
> fixed. In simple.tpl, this means doing the maximization w.r.t. x only.
> 2. Updating the value of the hyper-parameters, using the estimates of the
> random effects obtained in item 1.
>
> I think these values get set on lines 74 and 2707 of df1b2lap.cpp
> I've been trying to figure out if it makes sense to replace u.initialize()
> with initial_params::xinit(u); but after editing the file and reinstalling,
> the behaviour didn't change, so I wonder if I'm misunderstanding the
> installation instructions. I'll keep trying.
>
>
> Cheers,
> Mollie
> ------------------------
> Mollie Brooks, PhD
> Postdoctoral Researcher, Population Ecology Research Group
> http://www.popecol.org
> Institute of Evolutionary Biology & Environmental Studies, University of
> Zürich
>
>
>
> On 6 Nov 2013, at 1:31 PM, Ben Stevenson <bcs5 at st-andrews.ac.uk> wrote:
>
> Hi ADMBers,
>
> A while ago I noticed that all elements of a random_effects_vector seem to
> start out at 0, regardless of what appears in the .pin file. I have
> recently discovered that using the -noinit option results in behaviour that
> one would expect, i.e., random effects start values are read in from the
> .pin file.
>
> However, when you have a SEPARABLE_FUNCTION, this persists even with the
> -noinit option.
>
> Hans suggested a nice workaround, whereby you read in a vector of offsets
> (which can be set to your desired start values) from the .dat file, and use
> v as your random effect, where
>
> v = u + offset;
>
> and u is the random_effects_vector which has elements with start value 0.
>
> Can anyone shed any light on why start values in the .pin file for a
> random_effects_vector do seem to do what one would expect? The ADMB-RE
> manual suggests the .pin file initialisation should work.
>
> I have attached a minimum working example of the behaviour I am seeing.
> The compiled executable won't optimise anything, but start values for the
> random effects are printed out. I get a bunch of 0s, even though the .pin
> file provides the start values {1, 2, 3, 4, 5}.
>
> Cheers,
>
> Ben
> <test.dat><test.pin><test.tpl>
> _______________________________________________
> 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/20131106/f6adc905/attachment.html>