[Developers] Making changes to ADMB

Mark Maunder mmaunder at iattc.org
Wed Sep 12 11:31:21 PDT 2012


Trevor,

Cole could do this himself by getting access to the ADMB code from Johnoel, but it might take some learning, or he could suggest it to the developers and put it on the Redmine development website. We encourage people to get involved in the development of ADMB, so perhaps he could team up with one of the developers to implement it. I have copied the developers on this e-mail.

Mark

-----Original Message-----
From: Trevor A Branch [mailto:tbranch at uw.edu] 
Sent: Wednesday, September 12, 2012 11:17 AM
To: Mark Maunder; Cole Monnahan; Ian J. Stewart; Ian Taylor; James Thorson
Subject: Making changes to ADMB

Mark:

How do people go about making changes to ADMB itself? I ask because a quantitatively sharp student of mine, Cole Monnahan, worked over the summer with Ian Stewart and others at the NWFSC (Ian S is now at the IPHC), to improve the MCMC algorithm of ADMB.

In essence, he is able to take the variance-covariance matrix, read it into R, alter it, and write it back, in order to improve MCMC performance. As a proof of concept he has shown that replacing correlations greater than a particular value with a lower value, greatly improves MCMC convergence (by a factor of 10). In his example, he replaces a correlation of -0.94 with one of -0.5.

What I am thinking is that it would be amazingly useful to have a command line option in ADMB that would look like this:

model -mcmc 2000000 -mcsave 100 -nohess -noest -relaxed 0.5

This would replace all correlations with an absolute value greater than 0.5, with a value of 0.5. (i.e. -0.9 would become -0.5, and 0.9 would become 0.5).

I think that the R code worked out by Cole could be used as the basis for recoding ADMB in this way. I'll leave Cole to explain his method a little further.

Trevor

--
Assistant Professor, School of Aquatic and Fisheries Sciences, University of Washington, +1-206-221-0776 Organizer, Bevan Series on Sustainable Fisheries, Jan-Mar 2013 (Previous Bevan Series: http://courses.washington.edu/susfish/schedule.php)


More information about the Developers mailing list