[SciPy-dev] random seed in scipy.stats

[Robert Kern]

On Sat, Apr 19, 2008 at 5:48 PM, Hoyt Koepke <hoytak@gmail.com> wrote:
> Hello,
>
>  I may have missed something here, so I'll first ask: Is there a way to
>  specify a random number seed in scipy.stats?  I'm building a project
>  using it which requires exactly reproducible distribution samples, so
>  being able to specify a random seed is a must.
>
>  I've been looking around in the code, and I don't see a way to do it.
>  Even if it is there, it calls numpy.random directly, which would not
>  work reliably.

numpy.random.seed() is exposed, although I only recommend using it to
work around poorly designed APIs like the one in scipy.stats.

>  However, it seems that it'd be easy to add in by initializing a
>  RandomState object in the constructors of both rv_continuous and
>  rv_discrete, seeding it using an optional kw arg that defaults to
>  None.  All the code that currently draws samples from numpy.random
>  directly would then draw samples from this.  As far as I can see, this
>  wouldn't break backwards compatibility and would add in a useful
>  feature.  The only change (I think) would be to distributions.py.
>
>  Have I missed anything?  If this is acceptable, I'll gladly make the
>  change and post a diff and a unit test.

My preference is for a pattern like this:

def somefunc(x,y,z, prng=numpy.random):
    w = prng.standard_normal()
    ...

This allows the default uses to share the same generator state which
could avoid some problematic cases.

But yes, such a change would be very welcome. It might be a
substantial chunk of work, though, which is why I haven't done it.

-- 
Robert Kern

"I have come to believe that the whole world is an enigma, a harmless
enigma that is made terrible by our own mad attempt to interpret it as
though it had an underlying truth."
 -- Umberto Eco