[Numpy-discussion] Random number generators

Robert Kern robert.kern at gmail.com
Mon Sep 4 00:47:11 CDT 2006


Charles R Harris wrote:
> Hi Robert,
> 
> I am about to get started on some stuff for the random number generators 
> but thought I would run it by you first. I envisage the following:
> 
> uniform short_doubles -- doubles generated from a single 32 bit random 
> number (advantage: speed)
> uniform double, short_doubles on the interval (0,1) -- don't touch 
> singularities in functions like log (this is my preferred default)
> fast_normal -- ziggurat method using single 32 bit random numbers 
> (advantage: speed)
> fast_exponential -- ziggurat method using single 32 bit random numbers 
> (advantage: speed)
> MWC8222 random number generator (advantage: speed on some machines, 
> different from mtrand)
> 
> Except for the last, none conflict with current routines and can be 
> added without a branch. I expect adding MWC8222 might need more 
> extensive work and I will branch for that. So the questions are of 
> utility and naming. I see some utility for myself, otherwise I wouldn't 
> be considering doing the work. OTOH, I already have (C++) routines that 
> I use for these things, so a larger question might be if anyone else 
> sees a use for these. I like speed, but it is not always that important 
> in everyday apps.

I would prefer not to expand the API of numpy.random. If it weren't necessary 
for numpy to provide all of the capabilities that came with Numeric's 
RandomArray, I wouldn't want numpy.random in there at all.

Now, a very productive course of action would be to refactor numpy.random such 
that the distributions (the first four items on your list fall into this 
category) and the underlying PRNG (the fifth) are separated from one another 
such that they can be mixed and matched at runtime. A byproduct of this would 
expose the C API of both of these in order to be usable by other C extension 
modules, something that's been asked for about a dozen times now. The five items 
on your list could be implemented in an extension module distributed in scipy.

> I see that Pyrex is used for the interface, so I suppose that is one 
> more tool to become familiar with ;)

Possibly not. Pyrex was getting in the way of exposing a C API the last time I 
took a stab at it. A possibility that just occurred to me is to make an 
extension module that *only* exposes the C API and mtrand could be rewritten to 
use that API. Hmmm. I like that.

I can give some guidance about how to proceed and help you navigate the current 
code, but I'm afraid I don't have much time to actually code.

-- 
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





More information about the Numpy-discussion mailing list