BGonline.org Forums

41 opening RO w/XG 4 ply

Posted By: Nack Ballard
Date: Thursday, 26 November 2009, at 8:50 p.m.

In Response To: 41 opening RO w/XG 4 ply (neilkaz)

Thanks for the XG rollouts of opening 41 (money and GG). Here is the rollout data I have thus far:

Opening 41

Money
Snowie 3-ply [S $22] 31k
GNU 3-ply [S $9] 46k
XG 3-ply [S $8] 62k
XG 4-ply [S $21] 5k

DMP
Snowie 3-ply [S $4] 31k
GNU 3-ply [$ S6 U27 T45] 41k 5kUT [U S2 T12] 5k at -1-2pC

Gammon Save
Snowie 3-ply [S U20 $21] 31k

Gammon Go
Snowie 3-ply [S $8] 31k
GNU 3-ply [$ S8] 38k [$ S2] 12k at -4-2
XG 3-ply [$ S8] 3k

[My usual abbreviations apply: S = Split 24/23 13/9, $ = 13/9 6/5, U = 24/23 24/20, T = 13/8. The number following a letter/symbol is its error size in thousandths; e.g., "[S $8]" means that S beat $ by .008 in the rollout. The number after the brackets is the number of trials in thousands (usually a multiple of 1296), rounded down; e.g., 31k = 31,104. For scores other than money, Snowie's moves are not rolled out according to score. For GNU, I'm calling it's 2-ply "3-ply," equivalent depth to the other bots' 3-ply.]

For money, the XG 3-ply rollout is [S $8], agreeing with GNU, while the XG 4-ply rollout is [S $21], agreeing with Snowie. What’s going on here? In part, I suspect that the 4-ply rollout suffers from low variance (too few live cube trials, only 5k). If this file is ever extended, I predict the margin (currently at .021) will shrink to .015 or narrower. If the margin is still wide after extension, I will find the size of the difference between the XG 3-ply and XG 4-ply margins disturbing (frankly); indeed, until reconciled, it suggests that we can't really trust one of its ply settings, at least not for this type of position.

David Rockwell’s astute and well-articulated observations about GNU misplaying its replies to opening slots, which create a bias towards slotting in its rollout results, are consistent with my own observations comparing both Snowie second roll and and third roll positions with those of GNU. As you may recall, I characterized GNU (or, to be more precise, how its rollout results make it appear) as “slot-happy” in an e-mail last week.

Having said that, I notice in the reply rollouts (though with no predictable pattern detected thus far) that XG sometimes backs up GNU’s slot margins. An example is the identical [$ S8] results at GG above (for opening 41) – hope you eventually extend the 3024 trials.

Happy Thanksgiving.

Nack

Messages In This Thread

• 41 opening RO w/XG 4 ply
  neilkaz -- Wednesday, 25 November 2009, at 8:45 p.m.
  • 41 opening RO w/XG 4 ply
    christian munk-christensen -- Wednesday, 25 November 2009, at 9:05 p.m.
    • 41 opening RO w/XG 4 ply
      Timothy Chow -- Wednesday, 25 November 2009, at 10:08 p.m.
      • 41 opening RO w/XG 4 ply
        Rich Munitz -- Wednesday, 25 November 2009, at 11:26 p.m.
    • 41 opening RO w/XG 4 ply
      neilkaz -- Wednesday, 25 November 2009, at 11:16 p.m.
  • 41 opening RO w/XG 4 ply
    David Rockwell -- Wednesday, 25 November 2009, at 9:10 p.m.
  • 41 opening RO at GG w/XG 3ply
    neilkaz -- Wednesday, 25 November 2009, at 11:01 p.m.
    • I extended to 12960 games
      neilkaz -- Friday, 27 November 2009, at 3:56 p.m.
  • 41 opening RO w/XG 4 ply
    Nack Ballard -- Thursday, 26 November 2009, at 8:50 p.m.
    • 41 opening RO w/XG 4 ply
      neilkaz -- Thursday, 26 November 2009, at 9:40 p.m.
      • 41 opening RO w/XG 4 ply
        Nack Ballard -- Friday, 27 November 2009, at 12:47 a.m.
  • This is being extended to 20736 games
    neilkaz -- Friday, 27 November 2009, at 8:51 p.m.