[Python-Dev] Time for a change of random number generator?

[Chris Angelico]

On Fri, Feb 12, 2016 at 3:12 PM, Andrew Barnert via Python-Dev
<python-dev at python.org> wrote:
> On Thursday, February 11, 2016 7:20 PM, Stephen J. Turnbull <stephen at xemacs.org> wrote:
>
>
>
>> I think we should keep it around forever.  Even my slowest colleagues
>> are learning that they should record their seeds and PRNG algorithms
>> for reproducibility's sake. :-)
>
> +1
>
>> For that matter, restore Wichmann-Hill.
>
> So you can write code that works on 2.3 and 3.6, but not 3.5?
>
> I agree that it shouldn't have gone away, but I think it may be too late for adding it back to help too much.

You're probably right, but the point isn't to make the same code run,
necessarily. It's to make things verifiable. Suppose I do some
scientific research that involves a pseudo-random number component,
and I publish my results ("Monte Carlo analysis produced these
results, blah blah, using this seed, etc, etc"). If you want to come
back later and say "I think there was a bug in your code", you need to
be able to generate the exact same PRNG sequence. I published my
algorithm and my seed, so you should in theory be able to recreate
that sequence; but if you have to reimplement the same algorithm,
that's a lot of unnecessary work that could have been replaced with
"from random.deprecated_generators import WichmannHill as Random".
(Plus there's the whole question of "was your reimplemented PRNG
buggy" - or, for that matter, "was the original PRNG buggy". Using the
exact same code eliminates even that.)

So I'm +1 on keeping Mersenne Twister even after it's been replaced as
the default PRNG, -0 on reinstating something that hasn't been used in
well over a decade, and -1 on replacing MT today - I'm not seeing
strong arguments in favour of changing.

ChrisA