Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watnot!watmath!clyde!wayback!emh
From: emh@wayback.UUCP
Newsgroups: sci.math,sci.physics,sci.crypt,sci.math.stat
Subject: Re: Do you use RANDOM NUMBERS?
Message-ID: <1116@wayback.UUCP>
Date: Sat, 21-Mar-87 17:51:39 EST
Article-I.D.: wayback.1116
Posted: Sat Mar 21 17:51:39 1987
Date-Received: Sun, 22-Mar-87 17:53:42 EST
References: <5712@reed.UUCP> <553@entropy.ms.washington.edu> <5722@reed.UUCP>
Reply-To: emh@wayback.UUCP (E.M.Hummel)
Distribution: world
Organization: AT&T Bell Laboratories, Whippany NJ
Lines: 35
Keywords: Random numbers, leprechauns
Xref: utgpu sci.math:732 sci.physics:935 sci.crypt:252 sci.math.stat:77

In article <5722@reed.UUCP> mdr@reed.UUCP (Mike Rutenberg) writes:
>  "For most applications real random numbers are undesirable.  Consider
>   debugging a monte-carlo or simulation program when you cannot repeat
>   the sequence that led to the problem."  [Don Martin]

If a Monte Carlo simulation program is being tested, or is new
enough that there is a resonable probability of bug-problems, then
the sequence of random numbers can be stored on a peripheral.  In
this way, problems are reproducable and "true" random numbers can be
used.

>I agree that tracking a specific bug in a Monte-Carlo simulation is
>easier with a repeatable sequence, but what about running the thing for
>real.  Especially for long running simulations, I'm not sure how much I
>trust psudeo-random generator libraries, especially if I have a pretty
>good sense of just how random my hardware is.
>I find it a little curious that people are trying to simulate particles
>and other widgets using predictable perturbations (i.e. psudeo-random
>numbers) - it feels like playing with fire unless you really understand
>exactly how the simulation might be effected by the flavor of random numbers
>you're using -- AND THAT'S EXACTLY WHY YOU ARE DOING THE SIMULATION.
>Am I confused?

It depends on what is being modeled and how precise the model needs
to be specified.  Usually, the model used is approximate and the
level of approximation is known to make effects of correlations in
the pseudorandom number generator insignificant.  For Monte Carlo
experiments that study very subtle effects, the difference between
pseudo-randon, quasi-random and random can become significant.  For
these experiments, the accuracy and precision of the models used
must be very high.
-- 
Ed Hummel
AT&T Bell Laboratories
ihnp4!wayback!emh