Re: [Bug-gnubg] the sgf file is in the original posts (nt)

[Jim Segrave]

On Wed 14 Mar 2007 (13:10 +0100), Christian Anthon wrote:

So it is. I'm trying it now. I've already established the rollout
context from the .sgf file:

Cubeful, variance reduction, not an initial position, not truncate,
truncate at two sided database, truncate at one sided database,
Mersenne Twister, seed 1020115641

Player 0 2 ply Cube, deterministic, with pruning
       2 ply Cubeful, deterministic, with pruning
       keep 0 0 ply moves, add up to 8 within 0.16
              1 ply filter skipped

Player 1 is the same. This matches my normal settings (or close
enough) and I got 1.8 something JSDs.

OK - I can't reproduce this on an old version - pre-threading, nor can
I reproduce it on a very current version with the fixes for rankings
in the rollout windows.

I'd try building a version from mid-December, but it appears
savannah.gnu.org is currently unreachable.

You can easily calculate the JSD's if need be, although I agree it
might be nice to include it in the displayed information.

File->Export->Position as Text creates a text file with the data:

Rolled 21:*    1. Rollout          24/23 13/11                  Eq.: -0.159
       0.464 0.119 0.005 - 0.536 0.152 0.007 CL  -0.107 CF  -0.159
      [0.001 0.001 0.000 - 0.001 0.001 0.000 CL   0.002 CF   0.004]
        Full cubeful rollout with var.redn.
        10368 games, Mersenne Twister dice gen. with seed 1020115641
        and quasi-random dice
        Play: world class 2-ply cubeful prune [world class]
        keep the first 0 0-ply moves and up to 8 more moves within
        equity 0.16
        Skip pruning for 1-ply moves.
        Cube: 2-ply cubeful prune [world class]
     2. Rollout          6/3*                         Eq.:  -0.171 (-0.012)
       0.463 0.121 0.005 - 0.537 0.162 0.010 CL  -0.120 CF  -0.171
      [0.001 0.001 0.000 - 0.001 0.001 0.000 CL   0.002 CF   0.005]

Take the Standard Errors of the Cubeful equities for the two moves,
square them and sum the squares. Then take the square root of the
result. Divide the equity difference between the moves by that square
root and you have the JSD's:

sqrt(.004*.004 + .005*.005) = .006403

-0.012/.006403 = -1.87412, so it's 1.87 JSD's difference.

You mentioned a figure of about 5 JSDs. Now if you have the pre-25 Feb
2007 gnubg with the new threading code, AND when you select
Settings->Rollouts the panel shows cubeless, then those versions of
GNUBG would give the cubeless JSDs, in this case:

Equity difference is -0.13, the Standard Error for both is 0.002, so
-0.013/sqrt(.002*.002 + .002*.002) = -4.596 JSD's. It wouldn't
surprise me if this went up to around 5. In fact, changing my settings
to output 6 digits instead of 3 refined the figures and the cubeless
JSD is in fact 5.5268 JSD's. The rollout window uses the full
precision, so this is what I'd expect. 

I'm almost willing to bet that your default rollout settings when you
tried to extend this rollout were cubeless.



-- 
Jim Segrave           address@hidden