[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

Ultimate XG Settings – help please

Posted By: Taper_Mike
Date: Friday, 6 February 2015, at 8:51 p.m.

In Response To: Ultimate XG Settings – help please (Backgame)

I don’t have a monster like yours, so Neil, Igor, Stick, Ken Bame, or some of the others who traffic in extreme rollouts, may be a better source of info than me. Nevertheless, here is what I have gleaned over the years.

  1. Match evaluations – There are standards for Dual-Duel, BMAB, Mochy PR, Danes v World, etc., although I am not sure precisely what they are. I know, however, that Neil often uses stronger settings for his own matches than these groups use. In posts here, Neil has mentioned using XGR++ moves and cubes with a Gigantic move filter when he really gets serious.
  2. Rollouts – It probably depends on the position. Stick has argued convincingly that 3-ply/XGR (the standard XG settings) work as well as anything stronger for most positions. When you get a containment position or a difficult backgame, however, or one of the other types of positions where the bots are known to have trouble, you may want to revert to higher settings. In those cases, using a larger move filter is just as important as using more plys, XGR+, or XGR++.
     
    For the XG Opening Book, we used whatever we could get. When 4-ply (or better) rollouts were available, they were used. The rollouts I contributed use 3-ply/XGR. David Rockwell contributed the most rollouts. Early on, he standardized on 3-ply moves and cubes for his 2nd-roll positions, so much of the OB uses those settings. (For his recent work on 3rd-roll and later positions, David often uses 3-ply/XGR. For close positions, however, he is not averse to breaking out the 4-ply.)
     
    By convention, the number of trials is usually set to be a mulitple of 1296. Popular choices are 5k = 5184, 7k = 7776, 10k = 10368, 15k = 15552, 20k = 20736, 31k = 31104, 46k = 46656, and 62k = 62208. Anything less than 5k can be problematic. The top play may be identified with high confidence, but the exact margins by which it beats other plays will still have a lot of variance.
     
    When I release the next edition of my RolloutSummary spreadsheet, it will have more than 10,000 evals and rollouts of something like 3200 different positions. Consider what would happen if I accepted 95% confidence in a rollout. 5% of the time, the top play in a rollout would not be the best play. In my database, that would mean 5% of 10,000 rollouts would be wrong! Even at 99% confidence, my database would have 100 bad rollouts. As noted above, when the margin between plays is important, 100% confidence is not enough to shake out the variance. When I want to eliminate variance, I typically use 62k for plays within 0.01, 32k for plays within 0.02, 15k for plays within 0.03, 7k for plays within 0.04, and 5k for all other plays. Such large numbers are not necessary when all you want to do is determine the best play.
  3. Deep rollouts – Not sure about this one, but it may refer to using a large move filter.
  4. Bear-off database – I tried the largest for a while. That’s a database that goes to the midpoint. What I found is that a fast CPU is of little value. The larger databases are disk based. What you get, therefore, is a lot of disk thrashing. Rollouts are very slow. I ended up switching back to the standard database that stops at the 6pt. That’s entirely RAM based.

I look forward to hearing what others have to say, and welcome corrections to what I have written above. Drop me a line if you have not previously received my RolloutSummary and would like to be included on the distribution list.

Mike

Messages In This Thread

 

Post Response
Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:
 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.