Re: [Help-gsl] random number generators

[Toralf Niebuhr]

Am 18.04.2011 um 09:01 schrieb Heiko Bauke:

> Hi,
> 
> On Sat, 16 Apr 2011 23:01:52 +0200
> Toralf Niebuhr <address@hidden> wrote:
> 
>> we use random number generators for a simulation framework.
>> Because simulations should be reproducible, every time a simulation
>> is started, the generator is initialized with the same seed. Our
>> current setup the following:
>> 
>> first:
>> a feedback shift-register algorithm (started with the same seed for
>> every simulation and if needed explicitly set to a different seed)
>> then: a feedback shift-register algorithm for every random variable
>> initialized with a seed generated from the first generator.
>> 
>> The first generator is optimized to generate seeds, such that the
>> other generators are as much independent as possible.
>> 
>> What would be the optimal alternative, if we want to use the random
>> number generators for gsl? Speed is not to much important, statistic
>> qualities are important.
> 
> if quality matters, do not use feedback shift-register algorithms [1].
> gsl_rng_mt19937 and gsl_rng_cmrg (only 31 random bits) are not a bad
> choice.
> 
> 
>       Heiko
> 
> 
> [1] http://arxiv.org/abs/cond-mat/0307138

Question still is. If I have different random variables, should I use the same 
random number generator or should I have a new one for every random variable. 
If I should use a new one. How do I obtain the deterministic sequence of 
initial seeds such that the different generators produce good independent 
sequences?

Toralf Niebuhr