AMBER Archive (2009)

Subject: Re: [AMBER] use of random seed in multiple runs

From: Robert Duke (rduke_at_email.unc.edu)
Date: Fri Jun 05 2009 - 09:37:14 CDT

Hi Gustavo,
Well, I have not wanted to track through the code to figure out what would
really happen, though it would be simple enough to sample. BUT if it does
what you say here, that is potentially bad. The generator we use is pretty
much "stock" ranmar(), which is a pretty well-described and respected
generator with a very long period (one of several popularized by Marsaglia -
probably not the best available in its current incarnation, but not bad).
It is really two lagged fibonacci generators (with addition as the operation
on the fibonacci generator I believe), and these two generators are seeded
by another generator (a congruential generator of some sort). Various
additional monkey motion is applied to the two outputs of the LFG's to get
the final value in any one call. I WOULD stick to the recommended
constraints (seed > 0) unless the results of doing otherwise are tested (fed
through the diehard tests or some such - but that is work); my basic
philosophy on prng's is that taking liberties with them may produce
incorrect results which are hard to detect, so why risk increasing the
probability that you are generating junk if you don't have to. A dictum in
the MC physics community these days seems to be that it is a good idea to
run your experiments with two different prng's before you start to believe
the results.
Regards - Bob

----- Original Message -----
From: "Gustavo Seabra" <gustavo.seabra_at_gmail.com>
To: "AMBER Mailing List" <amber_at_ambermd.org>
Sent: Friday, June 05, 2009 9:28 AM
Subject: Re: [AMBER] use of random seed in multiple runs

On Thu, Jun 4, 2009 at 10:31 AM, Robert Duke<rduke_at_email.unc.edu> wrote:
> Well, I would not recommend doing that. It looks to me like there is
> nothing to stop one from doing this in the code, but the random number
> initialization code clearly states the seed should be > 0. I would have to
> dink with it to see what really happens when you do this.
> [...]

Bob, please correct me if I'm wrong here but, looking at the
random.fpp file in pmemd/src dir, it lloks like the seed provided by
the amber input is used to generate two internal seeds, is1 and is2:

=======================================================
  data is1max, is2max /31328, 30081/

! Construct two internal seeds from single unbound Amber seed:
!
! is1 and is2 are quotient and remainder of iseed/IS2MAX. We add
! one to keep zero and one results from both mapping to one.
! max and min functions keep is1 and is2 in required bounds.

  is1 = max((iseed / is2max) + 1, 1)
  is1 = min(is1, is1max)

  is2 = max(1, mod(iseed, is2max) + 1)
  is2 = min(is2, is2max)
========================================================

So, if you enter a seed of "-1", both is1 and is2 will end up being 1.
There's nothing "illegal" about that, but then you are again always
running your calculation with the same seeds.

Gustavo.

_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber

_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber