|
BGonline.org Forums
Bots and basic holding games....(again and long)
Posted By: neilkaz
Date: Sunday, 20 September 2009, at 3:39 p.m.
While going over the Depreli positions it became clear to me that a decent portion of the equity tossed by GNU and Snowie on cube decisions was due to overvaluing positions where the trailer has a single anchor game as the major source of his take equity. Both GNU (2 ply) and Snowie almost always and sometimes significantly overvalue the defender's chances in holding games and as a result both cube them later than is optimal and take cubes that they should clearly pass.
XG seems to suffer less from this disease from what I've seen. We'll let Frank post BGB evals here if he wants. Here's a basic 5 pt holding game that is a pass since the race lead is too much and there's decent timing. Posession of the bar pt means that a 65 will clear the mid pt later and that once the mid is cleared, future shot equity is small. Here's the GNU rollout. LOL ..no this position isn't TG, it is just that GNU in the rollout takes clear passes next turn.
The score (after 0 games) is: gnubg 0, user 0
Move number 8: user on roll, cube decision?
gnubg 146
user 123 Position ID: sG2LAQOYu4MHAA Match ID: cAkAAAAAAAAA
• user doubles
Alert: wrong double ( -0.0062)!
Cube decision Rollout cubeless equity +0.5885 Cubeful equities: 1. No double +1.0062 2. Double, take +1.0999 +0.0937 3. Double, pass +1.0000 -0.0062 Proper cube action: Too good to double, pass (6.2%) Rollout details
Win W g W bg Lose L g L bg Cubeless Cubeful Centered 1-cube 0.7882 0.0201 0.0007 - 0.2118 0.0085 0.0002 +0.5885 +1.0062 Standard error 0.0009 0.0007 0.0004 - 0.0009 0.0004 0.0001 0.0020 0.0047 Player gnubg owns 2-cube 0.7978 0.0205 0.0006 - 0.2022 0.0089 0.0005 +1.2147 +1.0999 Standard error 0.0009 0.0009 0.0004 - 0.0009 0.0006 0.0002 0.0043 0.0052 Full cubeful rollout with var.redn. 1296 games, Mersenne Twister dice gen. with seed 865987911 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]
Now for GNU 2 ply eval..which underestimates the leader's chances just as Snowie does. Snowie says 76.7% total wins and a .925 take on 3 ply.
The score (after 0 games) is: gnubg 0, user 0
Move number 8: user on roll, cube decision?
gnubg 146
user 123 Position ID: sG2LAQOYu4MHAA Match ID: cAkAAAAAAAAA
• user doubles
Cube decision 2-ply cubeless equity +0.5451 0.7667 0.0215 0.0002 - 0.2333 0.0101 0.0001 Cubeful equities: 1. Double, take +0.9316 2. Double, pass +1.0000 +0.0684 3. No double +0.8848 -0.0467 Proper cube action: Double, take
Now for GNU 3 ply, which lucks into the proper decision since 3 ply evals usually have the side on roll..ie the leader in cube decisions winning a couple percent more games.
The score (after 0 games) is: gnubg 0, user 0
Move number 8: user on roll, cube decision?
gnubg 146
user 123 Position ID: sG2LAQOYu4MHAA Match ID: cAkAAAAAAAAA
• user doubles
Cube decision 3-ply cubeless equity +0.5931 0.7880 0.0276 0.0004 - 0.2120 0.0108 0.0000 Cubeful equities: 1. Double, pass +1.0000 2. Double, take +1.0420 +0.0420 3. No double +0.9538 -0.0462 Proper cube action: Double, pass
OK here's a long XG rollout to compare and we see that it is close to the GNU rollout.
is Player 2
score: 0
pip: 146Money session pip: 123
score: 0
is Player 1XGID=----BbCCC---bD-abbbbb-----:0:0:1:D:0:0:3:0:10 double to 2 take ?
Analyzed in Rollout Player Winning Chances: 79.43% (G: 2.53% B: 0.08%) Opponent Winning Chances: 20.57% (G: 0.75% B: 0.02%) Cubeless Equities No Double: +0.589 Double: +1.214 Cubeful Equities No Double: +0.986 (-0.014) Double/Take: +1.117 (+0.117) Double/Drop: +1.000 Best Cube action: Double / Drop Rollout details 5184 Games rolled with Variance Reduction.
Moves and cube decisions: 3 plyConfidence No Double: ± 0.005 (+0.981...+0.991) Confidence Double: ± 0.007 (+1.110...+1.125) Double Decision confidence: 100.0% Take Decision confidence: 100.0% Duration: 5 hours 19 minutes 41 seconds Version: 1.03
Now for the XG 3 ply (equiv GNU 2 ply) evaluation. There's small fluctuation between XG's plies but XG is passing this cube on all plies.
is Player 2
score: 0
pip: 146Money session pip: 123
score: 0
is Player 1XGID=----BbCCC---bD-abbbbb-----:0:0:1:D:0:0:3:0:10 double to 2 take ?
Analyzed in 3 ply Player Winning Chances: 78.56% (G: 2.42% B: 0.06%) Opponent Winning Chances: 21.44% (G: 0.81% B: 0.02%) Cubeless Equities No Double: +0.571 Double: +1.176 Cubeful Equities No Double: +0.945 (-0.055) Double/Take: +1.025 (+0.025) Double/Drop: +1.000 Best Cube action: Double / Drop Version: 1.03
is Player 2
score: 0
pip: 146Money session pip: 123
score: 0
is Player 1XGID=----BbCCC---bD-abbbbb-----:0:0:1:D:0:0:3:0:10 double to 2 take ?
Analyzed in 4 ply Player Winning Chances: 79.38% (G: 2.62% B: 0.04%) Opponent Winning Chances: 20.62% (G: 0.90% B: 0.02%) Cubeless Equities No Double: +0.588 Double: +1.210 Cubeful Equities No Double: +0.963 (-0.037) Double/Take: +1.059 (+0.059) Double/Drop: +1.000 Best Cube action: Double / Drop Version: 1.03
I sure do hope that if and when (lets hope it is when) GNU gets additional training, that the result will have improved evaluation of basic holding games.
I posted this long disertation here since I am often asked this online and now can refer people to here.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.