[Developers] ADMB - AutoDiff
mmaunder at iattc.org
Tue May 26 19:52:20 PDT 2009
I am not sure of the answer to your question. Perhaps one of the ADMB
development team might be able to answer it for you.
From: Scott @ Zaita [mailto:scott at zaita.com]
Sent: Tuesday, May 26, 2009 4:32 PM
To: Mark Maunder
Subject: ADMB - AutoDiff
-----BEGIN PGP SIGNED MESSAGE-----
I'm working on a spatial population model for Antarctic Toothfish
(Dunn, A., and Rasmussen, S. 2008. Development of a spatially explicit
age-structured statistical catch-at-age population dynamics model for
modeling movement of Antarctic toothfish in the Ross Sea. WGSAM-08/14)
The of the minimizer's I'd like to implement is AutoDiff.
I've worked with a similar product called BetaDiff which uses a library
called ADOL-C (http://www.math.tu-dresden.de/~adol-c/). ADOL-C uses a
built in type called a "dvar" that is unwinds all modifications or
references made to the variables to propose a better candidate for the
Looking at the examples for autoDiff / minimize / t.cpp I see that your
energy function uses dvariables as well. Am I right in assuming that
AutoDiff uses the same principles as ADOL-C?
If so, This presents a problem, in that our model does a large amount of
calculations and using ADOL-C we went from 3MBs of memory usage to 3GBs
by enabling this (which required significant code changes). Is AutoDiff
going to be using the same order of magnitude in memory requirements?
Scientific Programmer / Software Developer
This message should be digitally signed to
verify it's authenticity. If you would like
secure email communication please reply with
your attached public signature (GnuPG).
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (MingW32)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
-----END PGP SIGNATURE-----
More information about the Developers