| |
BGonline.org Forums
Depreli positions 1-100 (of the 626) for GNU 4 ply and a 0 ply trunc rollout
Posted By: neilkaz
Date: Wednesday, 16 September 2009, at 1:09 a.m.
MCG has compared some GNU plies and XG on these... First 100 positions:
Checker error: GNU 2: 1.642 GNU 3: .882 XG3: 1.170 XGR: .387
Cube error: GNU 2: .608 GNU 3: .326 XG3: .194 XGR: .158
Total error: GNU 2/2: 2.250 GNU 3/3: 1.208 GNU 3/2: 1.490 XG3: 1.364 XGR: 0.545
Here's what I get for GNU default 4 ply. Cube error: .670 Checker error .561 Total error for 4 ply: 1.231
For my trucated rollout setting I used 1296 games of 0 ply play and cube truncated at 6. Cube error still .670 but checker error a very nice .222 for a total error of .892.
MCG you may want to recheck pos 58 and 62 as GNU made a .062 wrong decision in both and your 2 ply is .062 less than my 4 ply, but rechecking on 2 ply the same wrong decision was made.
This set of positions had cubes where odd plies in GNU clearly did better. I checked some of these cubes by truncating at an odd ply and then the correct result was obtained.
On the some of the cube errors 4 ply was closer to making the correct decision than was 2 ply and the same for the mini-rollout which also was closer.
The mini-rollout clearly was quite good for checker play and I suspect that a shorter rollout also would be good, although I do recall a couple of close plays that would have had a wrong result after 324 but switched by 1296.
I expect to continue this project with the next 100 positions and will look at XG. If XG does this well with the next 100, I think we can safely say that it is certainly really strong and a very good option for students, both for strength and speed. When I do XG on the next 100, I will also do 4 ply.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.