I’m currently working on some algorithms for which I’ve found it very useful to be able to generate the same random numbers in Python (using NumPy’s numpy.random module), C++ (using the Boost Random Number Library) and MATLAB (using the rand function).
I use the random numbers in tests, where being able to generate the same random numbers everywhere has been very useful. It allows me to compare the output of my algorithms exactly. For example, I might implement an algorithm in Python with NumPy first, write some tests, and then rewrite it in C++ (for the cases where I need the extra speed).
Getting MATLAB and NumPy to generate the same random numbers is straight forward. It seems NumPy uses the Mersenne Twister algorithm. MATLAB can be configured to do the same. Specifically, the following code generates the exact same 10 numbers with NumPy 1.0.3:
numpy.random.seed(0); numpy.random.rand(10,1)
and MATLAB R2006b:
rand('twister',0); rand(10,1)
However, the Mersenne Twister combined with a uniform distribution generator from the Boost Random Number Library don’t generate this same sequence. A quick glance through the NumPy and Boost code seemed to indicate that they use the same magic constants. I used the following C++ code with Boost 1.34.0:
#include <boost/random.hpp>
boost::mt19937 rng;
rng.seed(0U);
boost::uniform_real<> uni_dist(0, 1);
boost::variate_generator<boost::mt19937&, boost::uniform_real<> > uni(rng, uni_dist);
I’m guessing that NumPy and Boost simply do slightly different things with the bits that come out of the MT algorithm to generate the respective uniform distributions.
To work around this problem, I cooked up a Boost.Python wrapper for the Boost Random Library classes so that I could generate exactly the same numbers in Python and C++. You can take a look at the code here (the tests are still a work in progress):
The interface exposed by this wrapper behaves more like the stuff in numpy.random, compared to the wrapper in pyboost, which could be considered a very literal translation. It probably also wouldn’t be too hard to wrap these classes with MEX for MATLAB, guaranteeing you the “same” random numbers across all three platforms.
Disclaimer: I am not an expert on anything related to random numbers (except making them up in my head).
