Interesting results.
1) Could you also look at
        
 
     
 | | White is Player 2
score: 0
pip: 24 | | 1 point match | pip: 318
score: 0
Blue is Player 1 |
|
| XGID=-a-----------------CCCBBB-:0:0:1:12:0:0:0:1:10 |
| Blue to play 12 |
| 1. | Rollout1 | 24/21 | eq: -1.000 |
| Player:
Opponent: | 0.11% (G:0.00% B:0.00%)
99.89% (G:99.89% B:99.43%) | Conf.: ± 0.006 (-1.005...-0.994)
Duration: 7 minutes 59 seconds |
|
| 2. | XG Roller++ | 24/23 19/17 | eq: -0.995 (+0.005) |
| Player:
Opponent: | 0.24% (G:0.00% B:0.00%)
99.76% (G:98.90% B:97.98%) | |
|
| 3. | XG Roller++ | 24/22 20/19 | eq: -0.998 (+0.002) |
| Player:
Opponent: | 0.08% (G:0.00% B:0.00%)
99.92% (G:99.12% B:97.67%) | |
|
| 4. | XG Roller++ | 21/19 20/19 | eq: -0.998 (+0.001) |
| Player:
Opponent: | 0.08% (G:0.00% B:0.00%)
99.92% (G:99.65% B:99.56%) | |
|
| 5. | XG Roller++ | 20/18 19/18 | eq: -0.999 (+0.001) |
| Player:
Opponent: | 0.07% (G:0.00% B:0.00%)
99.93% (G:97.22% B:95.95%) | |
|
| |
1 10368 Games rolled with Variance Reduction.
Moves: 3-ply, cube decisions: XG Roller
|
eXtreme Gammon Version: 2.10, MET: Kazaross XG2
One interesting aspect of that result is the confidence interval. It is wider than what is obtained without using Variance reduction. That can happen when the evaluations are clueless since it is just adding noise.
2)Variance reduction works better when the evaluation is better. Thus, I am surprised by the confidence interval width for the strongest BGBlitz rollout being wider than the one for the strongest XG rollout. +-.0255 vs +- .013. This is coupled with having a win percentage further from 50% which should also lower the size of the confidence interval.