From: Gaussians 
To: 
Subject: ADBuilder
Date: Friday, November 12, 1999 2:20 PM

Dave Fournier  has a product called AD model builder, 
which maximizes functions (among other things) using automatic derivatives.
(Automatic derivatives are analytically computed derivatives done
automatically:  the user only needs to enter the objective function, as in
the equivalent of the loglik() proc in Gauss' MAXLIK.)  Dave, Ken Benoit,
and I worked to get a few of my examples into his program (he did the
coding) as a test run.  The deal was that we'd report on whatever we
found.

ADbuilder is a product separate from Gauss, but we were able to pull his
compiled code in via the foreign language interface.  The results were
really impressive.  The speed was between 13 and 68 times faster than
Gauss' maxlik without analytical derivatives specified. Obviously, adding
analytical derivatives would make maxlik faster, but that would have been
extremely tedious in the examples we gave him (and I'd rather spend
computer time than human time of course).

Ken and I now plan to learn how to use his program ourselves.

Gary King

 ..............................................................
 . Gary King                  Center for Basic Research       .
 . http://GKing.Harvard.Edu     in the Social Sciences        .
 . mailto:King@Harvard.Edu    34 Kirkland Street, #2          .
 . Admin.Assistant: 495-9271  Harvard U, Cambridge, MA 02138  .
 . Harvard-MIT Data Center:                                   .
 .   495-4734; FAX: 496-5149  Department of Government        .
 . Admin.Assistant: 495-9271  Littauer Center North Yard #M37 .
 . Direct: (617) 495-2027     Harvard U, Cambridge, MA 02138  .
 ..............................................................