Fergus Henderson wrote:
>
> On 02-Feb-2000, Simon Peyton-Jones <[EMAIL PROTECTED]> wrote:
> > Guess what? I have been driven to the conclusion that
> > the Haskell 98 Random library contains a fundamental, but
> > easily fixed flaw. So I propose to declare a new 'typo'.
> > (Yes, I am all too aware of the danger of typo-creep.)
> >
> > The library is currently based on the RandomGen class,
> > whose signature is:
> >
> > class RandomGen g wher
> > next :: g -> (Int, g)
> > split :: g -> (g, g)
> >
> > The difficulty is that there is absolutely no robust way to
> > use such a generator to generate random numbers uniformly
> > distributed in a given range.
>
> I think that is not true.
>
> > Why not? Because there's no
> > clue as to the precise range of Ints produced by next.
>
> You can figure out the range as you go along, by seeing what `next'
> returns. This makes things a bit more complex that it would otherwise
> need to be, but it's still possible. See the code below.
I like the clever code. Unfortunately, it's not just more complex, it's
less efficient. You will consume at least 3 random numbers from `next`
each time you need one new random number. That cuts the RNG's period by
at least two thirds.
I also think it could/would distort the statistical characteristics of
the RNG. I came up with a different hack that would have worked great
if the RNG was perfect, but as it was, I found it really skewed the mean
of output numbers. From that experience, I feel like your code would
have some similar effects. For instance, initial ascending sequences
from `next` get dropped by getRands:
[1,3,7,10,21,2] -> [2]
Same with descending sequences. That may not seem bad, but getRands is
called each time the user program requests a new number (via randN or
randsN). I think that means you'll see longer ascending or descending
sequences much less commonly from randN than you do from next.
You might understand these issues much better than I do; I'm not a
mathematician or statistician. I just want to point out the problems I
see.
---- cool hack snipped ----
Thanks,
Matt
