[Developers] MCMC in ADMB

[John Sibert]

Dear Cole, Melissa and Peter,

There was a presentation on this work at the June 2013 workshop in 
Seattle, anx I have been wondering what developed. This looks like a 
nice piece of work. I encourage you to get started making the proposed 
changes to the code and to post the MCMC guide on the website. It might 
also be possible to include the guide in the ADMB manual as a chapter.

Thanks for your efforts.
Cheers,
John

John Sibert
Emeritus Researcher, SOEST
University of Hawaii at Manoa
Honolulu HI (GMT-10)
808-294-3842 (mobile)

Visit the ADMB project http://admb-project.org/

On 11/12/2014 01:01 PM, Cole Monnahan wrote:
> Hi ADMB developers,
>
> We're grad students at the University of Washington and we've been
> looking through the source code to better understand how MCMC is
> implemented in ADMB. I started doing this because I wanted to specify
> an arbitrary covariance matrix, but Jim Ianelli encouraged me to
> explore some of the algorithms for the other available options. I
> couldn't find documentation for them, so I began writing them up with
> the idea of putting them into the ADMB manual. As time went on it grew
> more from a technical description to a starting place for users who
> had never used ADMB for Bayesian analyses before (as we were). So
> we're calling it a "guide" for lack of a better term. Our intent, as
> recommended by Jim, is to put it on the ADMB website, probably in a
> similar place to my other short write-up about the bounding functions
> <http://www.admb-project.org/examples/admb-tricks/covariance-calculations>.
>
> In doing this work we developed a list of updates we think would be
> worthwhile to include in future versions. We're not experienced C++
> programmers, but we would be happy to work with someone to help
> implement those into the source.
>
> Some proposed ideas for changes:
>
>   * Set mcsave default to 1, current default is to save nothing which
>     would be unexpected for a lot of users.
>   * Warn users if mcsave is invoked with the hybrid method (it has no
>     effect).
>   * Create new output file similar to .hst that outputs useful
>     information of the analysis (mcmc options used, time started,
>     ended, acceptance ratios, final scaling factor, etc.).
>   * Another file similiar to the .psv file (which contains only
>     parameters) that stores the draws of sd_report variables.
>   * Remove need for a sdreport variable for MCMC (checked with Dave,
>     not needed).
>   * Use a random seed if none is provided by user, or at least warn
>     user of the default which is to use the same seed value (36519).
>     In practice most users are reusing the same seed across models --
>     maybe not a good idea except for reproducibility which could be
>     done by looking at the seed in the .hst file.
>   * Clean up the console output for both Metropolis and hybrid runs.
>
> So please have a quick look through and let us know if (1) you find
> anything incorrect, and (2) whether you think it should be added into
> the ADMB manual or left as a separate document.
>
> Thanks,
> Cole, Melissa, and Peter
>
>
> _______________________________________________
> Developers mailing list
> Developers at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/developers