|
| From: | Ned Cross |
| Subject: | Re: [Bug-gnubg] Bug in analysis - position enclosed - more details |
| Date: | Mon, 28 Oct 2002 14:22:39 -0800 |
|
This must be one of those positions that
gives the "crashed" net trouble, but is there anything to be done about the
discontinuity in the analysis between the end of one move and the beginning of
the next?
Here's more data on the position. I can
send the .sgf if desired.
Ned
__________________________
Initial analysis parameters were: 2-ply, 16
cand, 0.360 tol.
Re-running analysis on world class ++ gives
same result:
1. Cubeful
2-ply 24/19
5/1*
Eq.: +0.827
0.774 0.000 0.000 - 0.226 0.067 0.004 2-ply cubeful 100% speed, 16 cand., 1 tol. [world class++] 2. Cubeful 2-ply 6/1* 5/1 Eq.: +0.689 ( -0.138) 0.730 0.000 0.000 - 0.270 0.012 0.000 2-ply cubeful 100% speed, 16 cand., 1 tol. [world class++] quick rollout gives:
1.
Rollout 6/1*
5/1
Eq.: +1.225
0.808 0.000 0.000 - 0.192 0.002 0.000 CL +0.614 CF +1.225 [0.003 0.000 0.000 - 0.003 0.009 0.001 CL 0.006 CF +0.021] 2.
Rollout 24/19
5/1*
Eq.: +0.854 ( -0.371)
0.842 0.000 0.000 - 0.158 0.051 0.003 CL +0.633 CF +0.854 [0.001 0.000 0.000 - 0.001 0.001 0.000 CL 0.003 CF +0.008] Full cubeful rollout with var.redn., with rotate rolls 1296 games, seed 1150722176 using Mersenne Twister Play and cube: 0-ply cubeful [expert] *****rollout seems to confirm the bots
"after-the-play" evaluation that the closeout wins >80% games. The
rollout shows the closeout to win fewer games than the alternate play (hence
lower cubeless equity), but gives a cubeful equity of greater than +1.000.
This is confusing, since GNU then thinks the proper cube action is double/pass,
so I'd expect the cubeful equity to be reported as +1.000
playing the alternate move (24/19 5/1*)
gives:
1. Cubeful 2-ply
24/19
5/1*
Eq.: +0.827
0.774 0.000 0.000 - 0.226 0.067 0.004 2-ply cubeful 100% speed, 16 cand., 0.36 tol. 2. Cubeful 1-ply 24/15 Eq.: +0.341 ( -0.486) 0.621 0.000 0.000 - 0.379 0.043 0.001 1-ply cubeful, 16 cand., 0.36 tol. 3. Cubeful 0-ply 24/19 12/8 Eq.: +0.359 ( -0.468) 0.672 0.000 0.000 - 0.328 0.083 0.002 0-ply cubeful [expert] 4. Cubeful 0-ply 6/1* 5/1 Eq.: +0.354 ( -0.473) 0.636 0.000 0.000 - 0.364 0.022 0.001 0-ply cubeful [expert] *****I interpret this to mean that GNU's
static evaluation has trouble with this position since it doesn't "see" the 6/1*
5/1 move, thinking it's worse than -0.360 from the "best"
move.
quick rollout of cube action after closeout
gives:
Cube analysis Rollout cubeless equity +0.682 (Money: +0.683) 1. Double,
pass +1.000
2. Double, take +1.220 ( +0.179) 3. No double +1.041 ( +0.000) Proper cube action: Double, pass Rollout details: Centered 1-cube: 0.843 0.000 0.000 - 0.157 0.003 0.000 CL +0.682 CF +1.041 [0.002 0.000 0.000 - 0.002 0.000 0.000 CL 0.005 CF +0.016] Player GNUfish owns 2-cube: 0.812 0.000 0.000 - 0.188 0.002 0.000 CL +1.416 CF +1.220 [0.003 0.000 0.000 - 0.003 0.000 0.000 CL 0.146 CF +0.022] Full cubeful rollout with var.redn., with rotate rolls 1296 games, seed 1150722176 using Mersenne Twister Play and cube: 0-ply cubeful [expert] *****Another confusing result since nodouble is
reported with a higher equity than double/pass, yet GNU recommends
double/pass. I for one can't see holding the cube here, but waiting must
lead to situations where GNU thinks your opp gets to a bare take and you
eventually win 2 pts.
|
| [Prev in Thread] | Current Thread | [Next in Thread] |