Re: [boost] [Boost.Random]Comments/Questions

From: Neal Becker (ndbecker2_at_[hidden])
Date: 2006-05-24 21:56:03

Darren O'Driscoll wrote:

> Hi all,
> I'm a relatively recent discoveror of these libraries, and for the last
> few weeks I've been like a kid in a toy shop as I've worked through them,
> and discovered the joys of the shared_ptr,bind,lambda,function libraries,
> etc.
> 
> I've just arrived at the Boost.Random library, an area quite close to my
> heart as I make a lot of use of random deviate generators in my work in
> Financial Mathematics. I really like the set up and design of the
> libraries, but I have a few suggestions for additions which might be nice
> to have. Apologies in advance if these have already been discussed here, I
> haven't quite got the hang of searching through old posts in an efficient
> way.
> 
> I hope noone's switched off during my long pre-amble, on to the meat of my
> post:
> 
> 1.  In general I've noticed that deviates for specific distributions are
> generated in one of two ways - a) according to a special algorithm, which
> usually involves generating one or more uniforms and doing some optimised
> transforms, or b) by generating a (0,1) uniform, and inverting the
> Cumulative distribution function.
> 
> Method a), which seems to have been stuck to in Boost.Random is generally
> the more efficient of the two methods, but there are a number of cases
> where b) is necessary, specifically when using low-discrepancy sequences
> such as SOBOL as a base for the generation, of when using the stratified
> sampling technique in Monte Carlo.
> 
> Would it be beyond the scope of this library to provide this kind of
> support?
> 
> 2. Following on from the last point, it would be very useful if SOBOL
> sequence generation was included in the library.
> 
> 3. I've noticed that at least one of the generators (Normal) uses a
> sub-optimal generator. I've attached an alternative implementation which I
> knocked together this evening which in my tests is about 30% faster. Note
> this was heavily influenced by an algorithm described in "Numerical
> Recipies in C".
> 
> 4. The list of distributions where deviate generators are provided is
> quite short. As mentioned before, I'm quite interested in this area, and
> would be happy to help with algorithms for some other distributions,
> though sadly I would estimate that my C++ skills are still quite a lot
> below the median here.
> 
> All comments welcome,
> 
Have you looked at unuran?  Looks interesting:
http://statistik.wu-wien.ac.at/unuran/