BGonline.org Forums

XGR++ versus a 1k Rollout

Posted By: Taper_Mike
Date: Friday, 13 December 2013, at 11:23 p.m.

In Response To: 61P-33C-31 — No Jacoby, No Beavers — XGR++ Eval (Philippe Michel)

XGR++ versus a 1k Rollout

If you search for "XGRoller" (no spaces) in the XG Help, you will find this:

XGRoller++ Checker
360 games with VR
Truncated after 5 moves using 3-ply
First 2 decisions in 3-ply
Then 2-ply checker and 3-ply cube

XGRoller++ Cube
360 games with VR
Truncated after 7 moves using 3-ply
Stop when reaching 0.010 Confidence (minimum 180 games)
First 2 decisions in 3-ply
Then 2-ply checker and 3-ply cube

Philippe Michel understands better than I do the advantages and disadvantages of using a truncated rollout such as XGR++ over a full rollout. He is the programmer who developed the new neural nets for GnuBg Version 1.

That said, in most “ordinary” positions, including most early game postions, I would expect XGR++ to be just as good as a 1k rollout in determining the top play. In those same positions, I think XGR++ will be better than a 1k rollout at estimating the margins between plays. That’s because there is less variance in a trunc than there is in a short rollout.

In the post cited by Philippe, Xavier concludes that “a full 3-ply 1296 game rollout appears as strong as a XGR++ analyze (0.115 and 0.117 PR respectively).” This judgement is based solely on finding the best play. It does not concern itself with the accuracy of the margins between plays.

The reason that experts like David Rockwell, Neil Kazaross, Stick and others prefer longer rollouts is because they are trying get a more accurate assessment of the margins between plays.

Mike

Messages In This Thread

• Catchup #2 - 61P-33C-21?
  Stick -- Monday, 1 December 2008, at 12:56 p.m.
  • Catchup #2 - 61P-33C-21?
    Coolrey -- Monday, 1 December 2008, at 1:01 p.m.
  • Catchup #2 - 61P-33C-21?
    Matt Cohn-Geier -- Monday, 1 December 2008, at 2:02 p.m.
    • 61P-33C-31
      onur alan -- Wednesday, 11 December 2013, at 7:41 p.m.
      • 61P-33C-31
        Taper_Mike -- Thursday, 12 December 2013, at 1:25 a.m.
      • 61P-33C-31 XG 1k RO
        onur alan -- Friday, 13 December 2013, at 1:45 p.m.
        • 61P-33C-31 — No Jacoby, No Beavers — XGR++ Eval
          Taper_Mike -- Friday, 13 December 2013, at 3:17 p.m.
          • 61P-33C-31 — No Jacoby, No Beavers — XGR++ Eval
            onur alan -- Friday, 13 December 2013, at 4:08 p.m.
            • 61P-33C-31 — No Jacoby, No Beavers — XGR++ Eval
              Philippe Michel -- Friday, 13 December 2013, at 7:26 p.m.
              • XGR++ versus a 1k Rollout
                Taper_Mike -- Friday, 13 December 2013, at 11:23 p.m.
                • XGR++ versus a 1k Rollout
                  onur alan -- Saturday, 14 December 2013, at 8:42 a.m.
                  • XGR++ versus a 1k Rollout
                    Taper_Mike -- Saturday, 14 December 2013, at 11:02 a.m.
        • 61P-33C-31 XG 5k RO
          onur alan -- Saturday, 14 December 2013, at 11:27 a.m.
      • 61P-33C-31 GNU 7k RO
        Philippe Michel -- Friday, 13 December 2013, at 7:15 p.m.
  • Catchup #2 - 61P-33C-21?
    neilkaz -- Monday, 1 December 2008, at 7:56 p.m.
  • 61P-33C-21 — No Jacoby, No Beavers — 3-ply XG2 Rollout
    Taper_Mike -- Monday, 16 December 2013, at 7:15 a.m.