| |
BGonline.org Forums
Bug in Opening roll rollouts in eXtreme Gammon
Posted By: eXtreme Gammon
Date: Tuesday, 19 April 2011, at 1:31 a.m.
From http://www.extremegammon.com/openingROBug.asp
Hello Everyone,
Recently, while checking a rollout in version 2, I came up with strange results. After investigating the problem I realized the issue had to do with the opening book.
What I realized is that ALL opening roll rollouts made with XG do have a bug if you are using the Opening book. This bug is only limited to the opening position and all 2nd roll rollouts are not affected and neither are any other regular positions.
The bug creates a predictable bias to the analyze of a specific position. For money games, when using 3-ply Rollout (checker setting) the error can be up to 0.007 equity between move. Using 4-ply Rollout it can be up to 0.006. The only good news is that because the bug creates a predictable error it is possible to retrofit any rollout to get the proper result. I’ve made a tool that allows you to fix your opening rollouts. You can get it at http://www.extremegammon.com/fixOpeningROInstall.exe
I know many have dedicated long days (if not months) to do some of these rollouts and I do apologize to them in particular and the community at large as many have relied on these results.
For now, people making opening roll rollouts should disable the opening book. To do so navigate in “C:\program files\eXtreme Gammon\” and delete or rename the file “OpeningBook.db”
Fell free to contact us at contact@extremegammon.com if you have any questions.
Xavier Dufaure de Citres GameSite 2000 Ltd.
Technical details of the error
The bug was due to the calculation of Variance reduction. For instance for a 4-ply rollout: on any move the program make a 4-ply evaluation before the roll, and then a 3-ply evaluation of the resulting move with the dice rolled. For position in the opening book, the 4-ply evaluation returns the value in the opening move. The problem it creates is that the average of all 36 dice values (3-ply) does not match the evaluation of the initial position. The bias created is the difference between the opening book value and the 4-ply analyze
Example: http://www.bgonline.org/forums/webbbs_config.pl?frames;read=93527
Claude's RO gives [S=$ U19]. However the Opening book give an equity of 0.006, -0.006 and -0.008 for S, $ and U. A 4-ply analyze is -0.002, -0.012 and -0.014. This create a bias of 0.008, 0.006 and 0.006 for S, $ and U.
So the rollout should be [$ S2 U19]
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.