[R-pkg-devel] R-extension requirement about third-party random number generators (RNG)

[John Clarke]

Hi folks,

Reference from the manual:

   -

   Nor should the C++11 random number library be used nor any other
   third-party random number generators such as those in GSL.”


I am working at wrapping an existing statistical model with RCPP so it can
be used in R. The model is written in C++ and uses the well-known Mersenne
Twister RNG C++ class. We instantiate 4 instances of the RNG to keep the
random variates independent.

I understand from the r-release manual
<https://cran.r-project.org/doc/manuals/r-release/R-exts.html#Writing-portable-packages-1>
that
we should not use third-party RNGs and rather use R's interfaces to R's
internal random number generator. If this is a correct reading of the
requirement for an R package/extension, it will add quite a bit of
complexity to the wrapper because we'll have to probably pass in references
to something like *rstream* objects to ensure the streams/substreams remain
independent. I fear that it may also might negatively affect performance of
the model because (if I understand correctly) we need to maintain the state
of a single RNG engine and due to the nature of our model, we don't
pre-generate all the RNGs but rather generate just a few before switching
to another stream. This would mean get/setting the RNG state with every
single draw unless substantial changes were made to batch generate the
numbers a priori.

Before we tackle this problem, I want to confirm that it is indeed
forbidden to use the Mersenne Twister C++ class in a RCPP R package if we
want to publish it on CRAN. Then, if it is a requirement, and assuming I
want to use the *rstream* RNG package, how can I instantiate an rstream
object in C++? I've succeeded at instantiating it in R and passing it
through, but this isn't ideal because it breaks encapsulation/ease-of-use
for the user.

Thanks,

-John

	[[alternative HTML version deleted]]