🚀 go-pugleaf

Hello!

Is there a statistically acceptable way to seed a random number generator
in a computer program?  I'm using the following RNG (the multiplicative
congruential "minimal standard" published in Communications of the ACM
a while back):

   Z(i) = 7**5 Z(i-1) mod (2**31 - 1)
   Z(0) = seed

I've been using the program's process id as a seed, but
  (a) I'm not at all sure that's a good idea, and
  (b) Suddenly I need several RNGs simultaneously.

So, the question is, is there a statistically "good" way to seed a number
(say 10, for concreteness' sake) of RNGs, such that they do not act the same
each time the program is run, and such that each stream is statistically
independent of the others, without requiring any intervention from the user?

Also, I'm certainly open to arguments about why I should use a RNG other
than the one described above...

Many thanks in advance,

Joe
-- 
______________________________________________________________________________
 Joseph W. Lavinus, Virginia Tech   /o)\    email: lavinus@cstheory.cs.vt.edu
 1001 University City Blvd., #F-9   \(o/  Blacksburg, VA 24060 / 703-953-2678
 "Let me play with it first and I'll tell you what it is later" - Miles Davis

In article <2532@creatures.cs.vt.edu> lavinus@csgrad.cs.vt.edu (Joseph Lavinus) writes:
>Is there a statistically acceptable way to seed a random number generator
>in a computer program?  I'm using the following RNG (the multiplicative
>congruential "minimal standard" published in Communications of the ACM
>a while back):
>
>   Z(i) = 7**5 Z(i-1) mod (2**31 - 1)
>   Z(0) = seed
>
>I've been using the program's process id as a seed, but
>  (a) I'm not at all sure that's a good idea, and
>  (b) Suddenly I need several RNGs simultaneously.
>
>So, the question is, is there a statistically "good" way to seed a number
>(say 10, for concreteness' sake) of RNGs, such that they do not act the same
>each time the program is run, and such that each stream is statistically
>independent of the others, without requiring any intervention from the user?

 As regards (a), the process id may be a perfectly reasonable seed, but one
 of the traditional seeds programmers use is some sort of system call which
 returns a time. On the Macintosh I use a call which gives me the number of
 seconds since midnight, Jan 1, 1900. Other systems allows the number of
 milliseconds since last boot. They're a good bet. Essentially, the problem
 is that a good way to seed is with a random number, which is why you wanted
 a random number generator in the first place ...

 For (b), why do you feel the need for a number of RNGs? Since each result
 taken from the stream is independent of the others (for a good RNG), just
 use one *good* RNG and take all your values from it. If you use several
 RNGs all of the same type, you may end up with highly dependent series
 (Imagine the disaster if they got seeded with values close to each other
 on the cycle; you'd get a perfect lag relationship)

>Also, I'm certainly open to arguments about why I should use a RNG other
>than the one described above...

 A reasonable refernce to examine is "Stochastic Simulation"; B.D. Ripley
 It discusses a number of generators and is fairly application-oriented.

-Graham Wills
 Trinity College Dublin, Ireland