Re: [boost] Additions to boost::random

From: Jens Maurer (Jens.Maurer_at_[hidden])
Date: 2002-02-05 15:59:42

Matthias Troyer wrote:
> The problem we face at the moment is that the arguments to the seed(...)
> function and to the non-default constructor are dependent on the
> generator.

I agree this is a non-orthogonality in the current interface that
makes it difficult to change generators by means of a single
"typedef" or equivalent.

> One solution would be to always require a seed(uint32_t ). However the
> problem there is that for generators like the LCG many integers are bad
> seed values, and the good/bad values differ from generator to generator.

Indeed.

> Also, a lagged Fibonacci generator, or a Mersenne Twister (MT) needs a
> long
> seed block instead of a single integer.

You could initialize a MT by means of a seed LCG, which itself
is seeded by a uint32_t.

> There are however solutions to this problem, as implemented e.g. in the
> SPRNG library (http://sprng.cs.fsu.edu/). These solutions take a
> "sequence number", a nonnegative integer, and convert it into a good
> seed for lagged Fibonacci generator. Similar algorithms can be
> implemented
> for the MT, or also for LCGs. This makes it possible to write more
> generic
> codes.

If algorithms that derive provably good seeds from a sequence number
are indeed available for a wide variety of random number generators,
the proposed additional interface appears to be reasonable, and even
better than the current state-based interfaces.

> Since I do not want to change the semantics of seed(uint32_t) where
> it is avaiable, I propose to add the following members to the generators:
> 
>    sequence(uint32_t n) ; // seeds the generator with the n-th good seed
>    static const uint32_t max_sequence; // n can take all values
> 0..max_sequence

I'm slightly concerned about using "uint32_t" in a generic
interface.  History has proven time and again that having a
fixed-sized bit width in some interface gets you into trouble in
the long run.  (16 bit wchar_t anyone?)

Do we need constructors that take a sequence number instead of a
state-based seed as well?

(Please make sure to include "random" in all subject lines pertaining
to the topic.  Thanks.)

Jens Maurer