Re: PLEASE READ: use of random

[Rick Riolo]

a. Our most heavy RN-using model sends an average of
   20M-50M messages to get numbers from uniform[0,1)
   for a single run.

b. The largest? Who knows---1 billion maybe?!

c. Well, I guess it depends on what you mean by
   an experiment.   A typical "experiment" for me
   is composed of a bunch of "cases" (unique set
   of parameter values), each case having a bunch
   of runs with different RN seeds.
   Maybe an average experiment has 15-20 cases,
   each with 20-40 runs.

d. no objection.

I think our biggest concern is that we still be able
to seed the RNGs with one (or maybe a couple) of seeds
to get a run started.    If you need store/restore
more state to checkpoint/continue a run, that's ok.

Our second concern is that the RNG's be fast.
We have even wondered if it would be useful to have
macros as the "interfaces" for getting random numbers.
Then it would be easy to replace a macro definition that uses
message sending to objects with one that uses
function calls, without making any changes to
the model codes at all.    (I'm assuming there is more
overhead in message passing than in function calls,
though I don't have good numbers for how much.
But if its say 10-20%, and if for "lightweight"
agents that are spending a lot of time just
getting RN's, that could be a good fraction of the run time.)

 - r

Rick Riolo                       address@hidden
Program for Study of Complex Systems (PSCS)
1061 Randall Lab     University of Michigan
Ann Arbor MI 48109-1120
http://pscs.physics.lsa.umich.edu/PEOPLE/rlr-home.html

On Fri, 21 Feb 1997, Sven N. Thommesen wrote:

> Date: Fri, 21 Feb 1997 13:46:49 -0600
> From: Sven N. Thommesen <address@hidden>
> To: address@hidden
> Subject: PLEASE READ: use of random
> 
>       
> After discussions with Barry at SwarmFest, I have concluded that something
> needs to be done with the way we use 'seeds' with the random generators.
> 
> To give me an idea of what to do / how to do it, I would be grateful if
> some of the heavy users out there (you know who you are!) would email me
> with answers to the following questions:
> 
> a) what is the *average* number of random variates needed for a single run
> of your model (keeping in mind that any call to a distribution that returns
> a double now uses 2 random variates);
> 
> b) what is the *largest* number of random variates you imagine needing for
> a single run?
> 
> c) how many independent runs do *you* consider sufficient for a single
> experiment?
> 
> d) I also have the following question for y'all regarding PMMLCG:
> 
> It has been an (implicit or explicit?) promise for the generators (except
> SWB) that you could save the latest output, and later feed it to the
> generator as a seed, and get the *same* sequence generated as if you had
> let it run on in the first place. I have discovered that this has not in
> fact been the case with PMMLCG! [ It delivers (state-1) as output, but
> seeds are fed directly into the state. We fix this by adding 1 to seeds
> before plugging them into the state. ]
> 
> So the question: if I fix this 'bug', the PMMLCG generators will be
> 'broken' in the sense that they'll behave differently than in the past.
> Does anyone have strong objections to my fixing this generator in this way?
> 
> 
> Sven Thommesen
> 
>