Re: Testing a class which utilizes randomness

jarod...

actually, to validate a PRNG is a very complex piece of sophisticated
statistical analysis.

as an example, let's take your observation. the numbers 1,2,3,4...n, served
in that order, will certainly pass your proposed bin for any number of bins
that's a divisor of n.

try knuth, volume 2, which dedicates over a third of its text to PRNG
algorithms and their analyses.

(a long and entirely irrelevant aside, cuz i've always been fascinated with
this stuff: if you have knuth, check out his algorithm A, p. 27. this is an
extremely fast (addition only) algorithm.

curiously, I've never seen an analysis of algorithm A that rejects it. but
I have seen that virtually *all* of the people who give their lives over to
such things consistently choose far more complex algorithms, and i'm
betting they have their reasons. i've used it several times, but because i
needed very long cycles. most default language libraries have PRNGs with
cycles in the low thousands, which is pathetic. algorithm A is 2^54th
periodicity.

if anyone knows of a serious analysis of A, i'd love to see it.

if a person wanted high-grade PRNG they'd need to research the alternatives
and pick a serious one. actual cryptologists use hardware RNG's rather than
algorithms. time periods between alpha particle reception is one way, and
there are others.)

seeya,
hill






On Fri, Apr 4, 2014 at 1:55 PM, Jarod Eells <[email protected]> wrote:

>
>
> On Fri, Apr 04, 2014 at 12:23:49PM -0400, Mark Levison wrote:
> > Desirable? - by definition - randomness can and should occasionally
> > duplicate something. Testing a random number generator is very hard.
> >
> > Cheers
> > Mark
>
> Testing a RNG doesn't have to be hard.
> Just have to roll the dice a sufficient number of times and bin out the
> results.
> Then check that the bins are filled according to the expected statistical
> distribution.
> If you want a higher accuracy for your test then make more bins and/or
> roll more numbers.
>
> Jarod Eells
>  
>