|
BGonline.org Forums
Search interval in XG version 1 and 2
Posted By: eXtreme Gammon
Date: Friday, 10 June 2011, at 2:56 p.m.
As I’ve been rechecking the search interval (SI) influence on version 2 engine. I though the results would of interest for the community.
I’ve made a study on the SI in the past http://www.extremegammon.com/SearchInterval.asp but this time I used a different methodology.
The first time, I had made the assumption that if a move was missed at low SI it was necessary worse than the one the larger SI made by using the 3-ply evaluation
This time around, rather than using the 3-ply evaluation, I used the set of 1000 money games that is almost completely rolled out. So it is possible that a RO reverse the 3-ply order abd that the normal SI picked the best move, while larger SI didn’t (by luck rather than skill).
Here is the data for 3-ply for the following SI : normal, Large, huge, gigantic and infinite. It shows the Elo increase compared to the normal SI for V1 (old study), V1 (new study) and V2 (new study). It also show the speed compared to the normal interval.
* Gigantic was not made in the previous study, the number shown here is an interpolation between Huge and infinite.
V1 old V1 New V2 New Speed Normal 0.00 0.00 0.00 1.00 Large 1.62 1.34 0.33 1.27 Huge 2.18 1.62 0.25 1.56 Gigantic 2.60* 1.98 0.24 2.52 Infinite** 2.99 2.33 0.20 5.11
** not available to userThis is Elo difference, to get the PR difference divide by 33. So in V2 using large SI gives you 0.01 PR
As I was a bit surprised that the dElo went down in V2 between large and Huge I check some of the difference. It turns out one decision in particular is the cause of that drop: XGID=aBcBBaB-----bC-A-bAd--AbA-:1:1:-1:54:240:222:3:0:10, normal and large play Bar/20 6/2, while huge and above play bar/20 8/4. In 3-ply it is 0.037 better, but after RO it turns out to be a blunder. With more game I am confident that the huge will be better than large, but even with 1000 games we are under the CI of the study
The new system (using Rollout) is more accurate and as expected the differences in V1 are lesser than in the previous study. In version 1 the effect of larger SI is rather small (0.05 PR between Normal and huge) and has a high cost (1.56 slower). In version 2 the effect of large SI is almost inexistent(0.01 PR between normal and large) I plan to check the SI influence on a XGR+, I'll put a follow up when i get the data.
Xavier
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.