I think the best way to solve this issue to not use a state at all. It is fast, reproducible even in parallel (if wanted), and doesn't suffer from the shared issue. Would be nice if numpy provided such a stateless RNG as implemented in Random123: www.deshawresearch.com/resources_random123.html


On Tue, May 12, 2015 at 2:18 PM, Neal Becker <ndbecker2@gmail.com> wrote:
In order to make sure all my random number generators have good
independence, it is a good practice to use a single shared instance (because
it is already known to have good properties).  A less-desirable alternative
is to used rng's seeded with different starting states - in this case the
independence properties are not generally known.

So I have some fairly deeply nested data structures (classes) that somewhere
contain a reference to a RandomState object.

I need to be able to clone these data structures, producing new independent
copies, but I want the RandomState part to be the shared, singleton rs

In python, no problem:

from numpy.random import RandomState

class shared_random_state (RandomState):
    def __init__ (self, rs):
        RandomState.__init__(self, rs)

    def __deepcopy__ (self, memo):
        return self

Now I can copy.deepcopy the data structures, but the randomstate part is
shared.  I just use

rs = shared_random_state (random.RandomState(0))

and provide this rs to all my other objects.  Pretty nice!

Those who fail to understand recursion are doomed to repeat it

NumPy-Discussion mailing list

ORNL/UT Center for Molecular Biophysics cmb.ornl.gov
865-241-1537, ORNL PO BOX 2008 MS6309