[Python-Dev] Re: Mersenne Twister

[Guido van Rossum]

> -- Write it in pure python as a drop-in replacement for Wichman-Hill.

Yup.  I think the seed arguments are different though -- MT takes a
single int, while whrandom takes three ints in range(256).

> -- Add a version number argument to Random() which defaults to two.
>     If set to one, use the old generator so that it is possible to recreate
>     sequences from earlier versions of Python.  Note, the code is much
>     shorter if we drop this requirement.  On the plus side, it gives more
>     than backwards compatability, it gives the ability to re-run a
>     simulation with another generator to assure that the result isn't
>     a fluke related to a generator design flaw.

I think this is useful.  But I'd like to hear what Tim has to say.

> -- Document David Abrahams's link to 
>     http://www.boost.org/boost/random/mersenne_twister.hpp as the
>     reference implementation and

Hm.  What part of that file contains the actual algorithm?  I gues the
function void mersenne_twister<DataType,n,m,r,a,u,s,b,t,c,l,val>::twist()
???

>     http://www.math.keio.ac.jp/matumoto/emt.html as a place for
>     more information.  Key-off of the MT19337 version as the most
>     recent stable evolution.

Sure.  It would be nice to have at least *some* documentation in-line
in case those links disappear.  Maybe you can quote the relevant C++
code from the Boost version (with attribution) in a comment.

> -- Move the existing in-module test-suite into a unittest.  Add a new,
>    separate unittest suite with tests specific to MT (recreating a few 
>    sequences produced by reference implementations) and with a battery
>    of Knuth style tests.  The validation results are at:  
>    http://www.math.keio.ac.jp/matumoto/ver991029.html

It might be fun to have some heavy duty tests (which take hours or
days to run) checked in but not run by default.  We usually do this by
not naming the test file test_foo.py; it can then be run manually.

> -- When we're done, have a python link put on the Mersenne Twister
>     Home Page (the second link above).

Sounds like they would be only too eager to comply. :-)

> -- Write, test and document the generator first.  Afterwards, explore
>     techniques for creating multiple independent streams:
>     http://www.math.h.kyoto-u.ac.jp/~matumoto/RAND/DC/dc.html

Isn't that trivial if you follow the WH implementation strategy which
stores all the state in a class instance?

--Guido van Rossum (home page: http://www.python.org/~guido/)