[ADMB Users] How to use the mcr command?

[dave fournier]

On 12-08-08 11:08 AM, Taylor, Nathan wrote:

Radford Neal  has a good introductory and more discussion of Hybrid MCMC.

The basic idea is this. You regard the value of the log-likelihood as a 
height function for
a surface. You pick a starting point and think of a ball at this point. 
You give the ball
a kick in some direction and it rolls around for a while with the motion
governed by a differential equation which you have to integrate. The two
parameters of the integration are the discrete step size and the number 
of steps for the
integration.  The main difference fromordinary MCMC
is that the LL corresponds to the height of the ball and
the total probability you use in the hybrid MCMC selection criterion is 
the height plus the kinetic energy.
That is the total energy or Hamiltonian.

Hbegin_Ham is the beginning value of the Hamiltonian and Ham is the 
value at the end of the path,
They are printed as a simple diagnostic to indicate if the step size is 
reasonable.

Since total energy  is conserved it is constant along the path (if the 
integration is exact.)

There are a few technical details to make the process reversible but 
that is about it.

On real problems the main technical things to tune are:

1.) How to use the Hessian.  If the problem is extremely badly conditioned
   one should probably learn how to play with the SVD.

2.) Picking step sizes and the number of steps.


I think the only way to get a handle on this is to work with the kinds 
of horrible models
we see in practice.

          Dave


> Inspired by your suggestion, I am just running hybrid mcmc on an
> assessment that I'm currently working on.  While it runs, what do the
> Hbegin-Ham and Ham fields refer to?
>
> Best
>
> Nathan
>
> -----Original Message-----
> From: users-bounces at admb-project.org
> [mailto:users-bounces at admb-project.org] On Behalf Of dave fournier
> Sent: August-08-12 7:53 AM
> To: users at admb-project.org
> Subject: Re: [ADMB Users] How to use the mcr command?
>
> It would be nice if you tried the hybrid mcmc to sse if it performs
> better. It needs exercising.
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users
>