[ADMB Users] hybrid mcmc

Ian Taylor Ian.Taylor at noaa.gov
Tue Aug 31 11:51:49 PDT 2010

  Hi Dave,
Thanks for the jump start on Hybrid MCMC.

I modified your objective function to something 
more complex and got the attached results 
comparing normal MCMC to the hybrid option. Run 
times for with hybrid MCMC is slower, but there's 
definitely a payoff in terms of more active 
sampling and far better exploration of the 
parameter space.

I have a couple questions for you:
First, it seems that within hybrid, ADMB 
automatically produces some output at every 
iteration. Piping to a file speeds things up, but 
is there a way to turn this extra output off or 
reduce the frequency? Actually, if I knew what the 
output was, maybe I'd be more excited about seeing it.

Second, is there any sort of rescaling of the jump 
function(s) in the hybrid MCMC, akin to what 
happens in the normal MCMC mode in ADMB? I suppose 
the epsilon and n-steps inputs are a manual 
analogue to the automatic rescaling, but I haven't 
really taken the time to study the theory behind 
hybrid MCMC and therefore don't fully understand 
how rescaling might work.

Thanks again for putting this into ADMB in the 
first place and letting us know how easy it is to 
use it.

On 8/31/2010 12:36 AM, dave fournier wrote:
> If anyone wants to play with it the hybrid mcmc 
> is already partially
> implemented.  What is missing is restarts, and 
> it currently only works
> for non random effects models.  Here is a 
> trivial example I call
> hysimple.tpl
>   int n
>  !! n=10;
>   init_vector x(1,n)
>   sdreport_number x1
>   objective_function_value f
>   x1=x(1);
>   f+=0.5*norm2(x);
>   if (mceval_phase())
>     cout << x1 << endl
> You do the hybrid mcmc with a command like
>  ./hysimple -hybrid -mcmc 2000 -hyeps .3 
> -hynsteps 20
> It does not create any hst files, just saves the 
> values in
> a psv file. You get the results with -mceval 
> afterwards.
>  ./hysimple -mceval
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: hybridMCMCtest.jpg
Type: image/jpeg
Size: 627881 bytes
Desc: not available
URL: <http://lists.admb-project.org/pipermail/users/attachments/20100831/15034684/attachment.jpg>

More information about the Users mailing list