Re: variance reduction [Was Re: [Bug-gnubg] Measuring performance levels]

[nis]

On 28 Oct 2002 at 18:48, Joern Thyssen wrote:

> I was just browsing the variance reduction code in gnubg. I didn't
> program it, I just saw that the variance reduction is performed
> with n-1 ply evaluations (since you perform 21 of those it's equivalent
> to a n ply evaluation). I was wondering whether this actually is
> necesary. Perhaps it should be an option, e.g., one can perform a 2-ply
> rollout but use 0-ply variance reduction.

I also tried browsing the code. Unfortunately the code is very far 
from easy to read for me (I guess it is a combination of me being 
unfamiliar with C, combined with the code being a little cryptic).

I believe a little higher level of comments in the code might help 
the casual browser ...

Anyway, I believe that I understand how this part of the code works 
now. I am also confident that we can use almost any settings for the 
evalcontext for Variance reduction without introducing bias. The 
question is of course: Will it give us better rollout results faster?

I can see now (in a rather blurry way) why Douglas thought of doing 2-
ply VR on a 4-ply rollout - or in general (n-2)-ply on n-ply. I 
believe that all (or most, depending on whether there is a difference 
in evaluation settings) positions needed for that evaluation will 
already be cached.
-- 
Nis Jorgensen
Van Bossestraat 29  1, 1051 JS Amsterdam
Home Mobile: +31 (6) 1496 3629
Work: +31 (20) 52 36 202