[ADMB Users] importance sampling sparse hessian
hskaug at gmail.com
Sun Feb 26 22:50:50 PST 2012
The importance sampling stuff was implemented
before the -shess (sparse matrix stuff), so it
probably never was implemented for -shess.
The question is whether it should. Importance
sampling is a more accurate integration technique,
but it does not work well in very high dimension (>1000),
which is exactly when you want to use -shess. Hence
I am tempted to say "no". Then this should be
clear from the docs, and there should also be
an error statement in the code. This seems
like an excellent job for the upcomming developers
Can you make a note in Redmine that this is an issue
that needs to be resolved one way or the other?
On Mon, Feb 27, 2012 at 12:39 AM, Mollie Brooks <mbrooks at ufl.edu> wrote:
> I ran a random effects model without importance sampling and it worked fine. The model uses separable functions and has a sparse hessian, so I was using -shess. When I added -is 100 I got the following output.
> mollie-brookss-macbook-4:testing molliebrooks$ ./sepvar_shts_sim -shess -ndi 50000 -is 100
> Hessian type 4
> inner maxg = -0.00030703 Inner second time = -0.00030703 Inner f = 21052.1
> f = 21052.11245390588 max g = 0.0003070296323131708
> Newton raphson 1 f = 21052.11245072833 max g = 5.794764150403482e-19
> lower index greater than upper index in dvar_vector:: dvar-vector(const predvar_vector&)
> I tracked this error to this function dvar_vector::dvar_vector(_CONST predvar_vector& pdv) in the file fvar_ar1.cpp.
> Then I tried running it without -shess just to see, and it worked. I'm not really familiar with the method behind importance sampling, but this shouldn't happen, right?
> Mollie Brooks
> Ph.D. Candidate
> NSF IGERT Fellow
> Biology Department
> University of Florida
> mbrooks at ufl.edu
> Users mailing list
> Users at admb-project.org
More information about the Users