[R] Random seed problem in MCMC coupling of chains

[Jan T. Kim]

On Wed, Jun 08, 2005 at 12:55:07PM -0400, Gabor Grothendieck wrote:
> That could be addressed like this (where changing the offset
> changes the experiment).
> 
> offset <- 123
> 
> niter <- 3
> nchain <- 2
> for (i in 1:niter) { # iterations
>  for (j in 1:nchain) { # chains
>    set.seed(i+offset)
>    a <- runif(1)
>    cat("iter:", i, "chain:", j, "runif:", a, "\n")
>  }
> }
> 
> On 6/8/05, Paul Gilbert <pgilbert at bank-banque-canada.ca> wrote:
> > Beware that your easy trick will give you the same result every time you
> > run it. You need a better scheme if you actually intend to get a new
> > experiment each time you run it.

That's not a bad thing per se; in fact, it's a good thing to be able
to exactly reproduce the results you obtained with your software.
Personally, I dislike the "convenience" feature of random number
generators of generating a seed, frequently based on the system time,
if none has been set explicitly; I always set a seed and frequently
make the seed a commandline option or part of the control parameter
file or the like.

>From this perspective, Gabor's solution seems perfect to me.

> > Paul
> > 
> > Gorjanc Gregor wrote:
> > 
> > > Thanks to Duncan, Dimitris as well as James for answers. I'll provide
> > > here also example from James, which seems to be the easiest of them
> > > all and was not posted to the list:
> > >
> > > niter <- 3
> > > nchain <- 2
> > > for (i in 1:niter) { # iterations
> > >   for (j in 1:nchain) { # chains
> > >     set.seed(i)
> > >     a <- runif(1)
> > >     cat("iter:", i, "chain:", j, "runif:", a, "\n")
> > >   }
> > > }
> > >
> > > Note that seed is set with iteration counter. This is really neat and
> > > simple. I am just wondering if this is OK from "RNG point of view". Can
> > > someone comment on that?

The only concern I could think about is the case of a bad random number
generator, in which the first couple of values are not entirely
uncorrelated to the seed. But I'd be very surprised if that was a
problem with R's RNGs -- I guess it's memories of lousy implementations
C library rand() functions that make me write this remark.

Best regards, Jan
-- 
 +- Jan T. Kim -------------------------------------------------------+
 |    *NEW*    email: jtk at cmp.uea.ac.uk                               |
 |    *NEW*    WWW:   http://www.cmp.uea.ac.uk/people/jtk             |
 *-----=<  hierarchical systems are for files, not for humans  >=-----*