R help - Jun 2005 - FW: Random seed problem in MCMC coupling of chains

And a last post from Paul Gilbert. Thanks to all! This disscusion was
really beneficial for me!

-----Original Message-----
From: Paul Gilbert [mailto:pgilbert at bank-banque-canada.ca]
Sent: sre 2005-06-08 21:01
To: Gorjanc Gregor
Subject: Re: [R] Random seed problem in MCMC coupling of chains
 
Gorjanc Gregor wrote:
> Thanks to Paul and Gabor for additional tips/examples. Actually, I find
> Pauls suggestion with setRNG also nice and is exactly what I wanted. 
> Paul, if I understand this correctly, your suggestion with setRNG does not
> alter "RNG flow", it just takes care that chains really have
equal seeds.
> I remember that I have read somewhere that destroying "RNG flow over
and
> over to get real randomness" is not a good idea. Can someone confirm
this?
In general it is a bad idea to make up your own scheme for setting or 
resetting the RNG. People put a lot of work into studying the properties 
of a RNG. When you mess with it then it is unclear what the result will 
be. It certainly won't be tested unless you test it yourself. If your 
intention is to do research on RNGs then you may want to do that, but if 
your intention is to do other research and just use the RNG, then don't 
mess with it by resetting it with your own scheme.

One additional thing you may want to do is record the initial setting of 
the RNG information so that you can reproduce the experiment if you want 
to (see modification below). The idea in setRNG is to not interfere with 
the flow, only add a few utilities to help record and reset everything 
when that is what is required.

In your example it is important that you generate the same number of 
random numbers in each pass through the chain. If that is not the case 
then even with the setRNG utilities there is a subtle change that you 
are introducing.

HTH,
Paul
> 
> niter <- 3
> nchain <- 2
   startingRNG <- setRNG()
> for (i in 1:niter) { # iterations
>   tmpSeed <- setRNG()
>   for (j in 1:nchain) { # chains
>     setRNG(tmpSeed)
>     a <- runif(1)
>     cat("iter:", i, "chain:", j, "runif:", a,
"\n")
>   }
> }
> 
> iter: 1 chain: 1 runif: 0.8160078
> iter: 1 chain: 2 runif: 0.8160078
> iter: 2 chain: 1 runif: 0.4909793
> iter: 2 chain: 2 runif: 0.4909793
> iter: 3 chain: 1 runif: 0.4425924
> iter: 3 chain: 2 runif: 0.4425924
> 
> [... removed other stuff ...]
> 
> Lep pozdrav / With regards,
>     Gregor Gorjanc
> 
> ----------------------------------------------------------------------
> University of Ljubljana
> Biotechnical Faculty        URI: http://www.bfro.uni-lj.si/MR/ggorjan
> Zootechnical Department     mail: gregor.gorjanc <at> bfro.uni-lj.si
> Groblje 3                   tel: +386 (0)1 72 17 861
> SI-1230 Domzale             fax: +386 (0)1 72 17 888
> Slovenia, Europe
> ----------------------------------------------------------------------
> "One must learn by doing the thing; for though you think you know it,
>  you have no certainty until you try." Sophocles ~ 450 B.C.
> ----------------------------------------------------------------------
> 
> 
> 
> 
> 
>

On 6/8/05, Gorjanc Gregor <Gregor.Gorjanc at bfro.uni-lj.si>
wrote:
> And a last post from Paul Gilbert. 
> 
> -----Original Message-----
> From: Paul Gilbert [mailto:pgilbert at bank-banque-canada.ca]
> 
> In your example it is important that you generate the same number of
> random numbers in each pass through the chain. If that is not the case
> then even with the setRNG utilities there is a subtle change that you
> are introducing.
> 
Note that this is actually one of the advantages of the last solution
I posted, namely the one with the chain list.  It maintains the flow
along each chain even if different chains use different number of calls
to the random number generator.