From: Mark Harris 
Subject: AD Model Builder 
Date: Thursday, October 28, 1999 5:36 AM

Hi Gaussians,


We took Dave Fournier up in his offer (see below) to provide code to 
perform nonlinear estimation procedures (constrained maximum likelihood) 
using his package AD Model Builder for a model we are working on.

This is a description of our experience with his software.

The model is a qualitative response discrete choice model, a combination of 
multinomial logit, DOGIT and OGEV models. It is nonlinear maximum 
likelihood problem, with typically 50-100 (or more) parameters. We had been 
estimating the parameters in GAUSS using CML.
This appeared to take a long time (ie a number of days) and we quite often 
had convergence problems (function failing, not appearing to reach "true" 
maximums etc.).

The AD Model Builder software was much faster (typically hours as opposed 
to days) and generally gave very similar results to CML. Moreover, Dave 
transcribed my Gauss code in a matter of days.

We have not yet tried the DLL option of producing a DLL with ADMB which can 
be called from GAUSS and so can't comment on that feature. (ADMB also can 
produce stand along executables).

We feel that ADMB is potentially a powerful tool to augment GAUSS for this 
kind of nonlinear estimation problem, especially when Gauss seems to be 
taking a long time to converge.


Mark Harris

 >   Hi Gaussians,
 >   AD Model Builder is a package for creating nonlinear statistical
 >   models. For the right kind of problem it is very fast. It would be
 >   a good substitute for C or FORTRAN which users often turn to when
 >   they wish to solve a computationally intensive nonlinear estimation
 >   problem. Development with ADMB is much faster and simpler than with
 >   C or FORTRAN. ADMB can create stand alone programs or routines which
 >   can be called from GAUSS (or Splus). The best kind of problem to
 >   test ADMB is a numerically difficult nonlinear problem with a
 >   lot (say 20-1000) of parameters.

 >   I am looking for a number of such problems to showcase my software.
 >   If you think you have such a problem please contact me and we'll
 >   see if it seems to be a good candidate.

 >   What you get -- well you get your problem solved and a free copy
 >   of my software.

 >   What I get -- I hope you would send a message to this list saying
 >   how (presumably) happy you are with my software as an extension
 >   To GAUSS.

 >              Cheers,
 >              Dave

 >Dave Fournier
 >Otter Research Ltd
 >PO Box 2040, Sidney B.C. V8L 3S3

Mark Harris, Senior Research Fellow
The Melbourne Institute of Applied Economic and Social Research
University of Melbourne, Vic. 3010, Australia
tel +61 3 9344 5354: fax +61 3 9344 5630