1. Forum
  2. USENET
  3. rec.games.backgammon

  • Surprised by XG

    From peps...@gmail.com@pepstein5@gmail.com to rec.games.backgammon on Sun Dec 10 04:44:09 2023
    From Newsgroup: rec.games.backgammon

    I'm very surprised by XG's play here -- 18/15 7/6 because it crunches on 33. Why not 18/14 which plays well on all next rolls?

    Thank You.

    Paul

    XGID=-BBB-CCa-----B----bbbbbbbA:1:-1:-1:11:2:8:0:11:10
    X:XG Roller+ O:Daniel

    Score is X:8 O:2 11 pt.(s) match.
    +13-14-15-16-17-18------19-20-21-22-23-24-+
    | X | | O O O O O |
    | | | O O O O O |
    | | | O O |
    | | | |
    | | | |
    | |BAR| |
    | | O | |
    | | | |
    | | | | +---+
    | O X | | X X X X X X | | 2 |
    | O X | | X X X X X X | +---+
    +12-11-10--9--8--7-------6--5--4--3--2--1-+
    Pip count X: 74 O: 96 X-O: 8-2/11
    Cube: 2, X own cube
    X to play 11

    1. XG Roller+ 18/15 7/6 eq:+1.076
    Player: 96.68% (G:22.96% B:0.06%)
    Opponent: 3.32% (G:0.01% B:0.00%)

    2. XG Roller+ 7/5(2) eq:+1.076
    Player: 96.85% (G:22.62% B:0.06%)
    Opponent: 3.15% (G:0.01% B:0.00%)

    3. XG Roller+ 18/14 eq:+1.076
    Player: 96.97% (G:22.19% B:0.06%)
    Opponent: 3.03% (G:0.00% B:0.00%)

    4. XG Roller+ 7/6 7/4 eq:+1.075 (-0.001)
    Player: 96.72% (G:22.79% B:0.06%)
    Opponent: 3.28% (G:0.01% B:0.00%)

    5. XG Roller+ 18/16 7/6(2) eq:+1.067 (-0.009)
    Player: 96.24% (G:23.03% B:0.06%)
    Opponent: 3.76% (G:0.01% B:0.00%)


    eXtreme Gammon Version: 2.10, MET: Kazaross XG2
    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From peps...@gmail.com@pepstein5@gmail.com to rec.games.backgammon on Sun Dec 10 05:09:51 2023
    From Newsgroup: rec.games.backgammon

    On Sunday, December 10, 2023 at 12:44:11rC>PM UTC, peps...@gmail.com wrote:
    I'm very surprised by XG's play here -- 18/15 7/6 because it crunches on 33. Why not 18/14 which plays well on all next rolls?
    Actually, the stats kind of answer my own question.
    My own play of 18/14 does actually win more games, and probably for the reason I said.
    But the extra games is only 0.3%. Does this look right? So the prob of 33 followed by me
    entering is 1%. So XG is right if I win around a third of the games when that happens.
    This looks totally reasonable.
    I can see how XG's play is more gammonish because it plays to 6 to avoid an inefficient bear-in.
    And apparently the extra gammons compensate. Instructive example of the difference between
    bots and strong players. I think Stick would play 18/14, as would any good human player.
    Paul

    Thank You.

    Paul

    XGID=-BBB-CCa-----B----bbbbbbbA:1:-1:-1:11:2:8:0:11:10
    X:XG Roller+ O:Daniel

    Score is X:8 O:2 11 pt.(s) match. +13-14-15-16-17-18------19-20-21-22-23-24-+
    | X | | O O O O O |
    | | | O O O O O |
    | | | O O |
    | | | |
    | | | |
    | |BAR| |
    | | O | |
    | | | |
    | | | | +---+
    | O X | | X X X X X X | | 2 |
    | O X | | X X X X X X | +---+
    +12-11-10--9--8--7-------6--5--4--3--2--1-+
    Pip count X: 74 O: 96 X-O: 8-2/11
    Cube: 2, X own cube
    X to play 11

    1. XG Roller+ 18/15 7/6 eq:+1.076
    Player: 96.68% (G:22.96% B:0.06%)
    Opponent: 3.32% (G:0.01% B:0.00%)

    2. XG Roller+ 7/5(2) eq:+1.076
    Player: 96.85% (G:22.62% B:0.06%)
    Opponent: 3.15% (G:0.01% B:0.00%)

    3. XG Roller+ 18/14 eq:+1.076
    Player: 96.97% (G:22.19% B:0.06%)
    Opponent: 3.03% (G:0.00% B:0.00%)

    4. XG Roller+ 7/6 7/4 eq:+1.075 (-0.001)
    Player: 96.72% (G:22.79% B:0.06%)
    Opponent: 3.28% (G:0.01% B:0.00%)

    5. XG Roller+ 18/16 7/6(2) eq:+1.067 (-0.009)
    Player: 96.24% (G:23.03% B:0.06%)
    Opponent: 3.76% (G:0.01% B:0.00%)


    eXtreme Gammon Version: 2.10, MET: Kazaross XG2
    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From MK@murat@compuplus.net to rec.games.backgammon on Mon Dec 11 02:46:29 2023
    From Newsgroup: rec.games.backgammon

    On December 10, 2023 at 6:09:53rC>AM UTC-7, peps...@gmail.com wrote:
    I'm very surprised by XG's play here --
    18/15 7/6 because it crunches on 33.
    I agree that 18/15 7/6 is not the best play.
    Why not 18/14 which plays well on all
    next rolls?
    I think Stick would play 18/14, as would
    any good human player.
    Not being one of them/you sick world-class
    or even merely good gamblegammon players,
    I can't speak for them but I wouldn't play that.
    I wouldn't argue that 18/14 is not better than
    18/15 7/6 but I don't see why the latter is the
    better play either.
    Stacking on high points is better for bearing
    off and there is no reason to hurry the back
    checker. Stacking on 5-point here is actually
    better than on 6-point. So, this is the perfect
    roll for 7/5(2) especially because 66 won't
    hurt as it will be unplayable and every other
    roll will play either better or not any worse
    after 7/5(2).
    MK
    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From Timothy Chow@tchow12000@yahoo.com to rec.games.backgammon on Tue Dec 12 23:27:44 2023
    From Newsgroup: rec.games.backgammon

    For what it's worth, below is an XGR+ rollout of the position. Murat's
    play is on top, albeit by a very narrow margin. Note that the numbers
    are displayed to 4 decimal places instead of the usual 3 decimal places.

    XGID=-BBB-CCa-----B----bbbbbbbA:1:-1:-1:11:2:8:0:11:10

    Score is X:8 O:2 11 pt.(s) match.
    +13-14-15-16-17-18------19-20-21-22-23-24-+
    | X | | O O O O O |
    | | | O O O O O |
    | | | O O |
    | | | |
    | | | |
    | |BAR| |
    | | O | |
    | | | |
    | | | | +---+
    | O X | | X X X X X X | | 2 |
    | O X | | X X X X X X | +---+
    +12-11-10--9--8--7-------6--5--4--3--2--1-+
    Pip count X: 74 O: 96 X-O: 8-2/11
    Cube: 2, X own cube
    X to play 11

    1. Rollout-| 7/5(2) eq:+1.0841
    Player: 97.30% (G:21.85% B:0.21%)
    Opponent: 2.70% (G:0.02% B:0.00%)
    Confidence: -#0.0010 (+1.0831..+1.0851) - [77.9%]

    2. Rollout-| 18/15 7/6 eq:+1.0835 (-0.0006)
    Player: 97.10% (G:22.33% B:0.21%)
    Opponent: 2.90% (G:0.01% B:0.00%)
    Confidence: -#0.0011 (+1.0824..+1.0845) - [18.6%]

    3. Rollout-| 18/14 eq:+1.0829 (-0.0011)
    Player: 97.35% (G:21.62% B:0.22%)
    Opponent: 2.65% (G:0.00% B:0.00%)
    Confidence: -#0.0010 (+1.0820..+1.0839) - [3.3%]

    4. Rollout-| 7/6 7/4 eq:+1.0823 (-0.0018)
    Player: 97.11% (G:22.16% B:0.21%)
    Opponent: 2.89% (G:0.02% B:0.00%)
    Confidence: -#0.0010 (+1.0812..+1.0833) - [0.2%]

    -| 2592 Games rolled with Variance Reduction.
    Dice Seed: 271828
    Moves and cube decisions: XG Roller+
    Search interval: Gigantic

    eXtreme Gammon Version: 2.19.211.pre-release, MET: Kazaross XG2

    ---
    Tim Chow
    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From MK@murat@compuplus.net to rec.games.backgammon on Tue Dec 12 23:55:00 2023
    From Newsgroup: rec.games.backgammon

    On December 12, 2023 at 9:27:48rC>PM UTC-7, Timothy Chow wrote:
    For what it's worth, below is an XGR+ rollout
    of the position. Murat's play is on top, albeit
    by a very narrow margin.
    X to play 11
    1. Rollout-| 7/5(2) eq:+1.0841
    2. Rollout-| 18/15 7/6 eq:+1.0835 (-0.0006)
    3. Rollout-| 18/14 eq:+1.0829 (-0.0011)
    4. Rollout-| 7/6 7/4 eq:+1.0823 (-0.0018)
    As plain analysis,
    2-ply plays really bad
    3-ply picks 1=18/14, 2=7/5(2), 3=18/15 7/6
    xg-roller picks the same as 3-ply
    4-ply picks 1=7/5(2), 2=18/14, 3=18/15 7/6
    xg-roller+ picks 1=18/15 7/6, 2=7/5(2), 3=18/14
    xg-roller++ picks 1=18/14, 2=18/15 7/6, 3=7/5(2)
    I find this pretty fucking funny. Except for 3-ply
    and xg-roller all other plies give every different
    conbination possible. :)
    Granted they are all very close but shouldn't we
    expect a little more consistency?
    So then, do you guys take xg-roller+ rollout more
    accurate than plain xg-roller++?
    I suppose an xg-roller++ rollout would take some
    hours but will it agree with the xg-roller+ rollout?
    I personally wouldn't bet on it. So, what gives..?
    Interestingly Noo-BG's various ply analyses are
    just as inconsistent but then even the 0-ply with
    maximum noise rollout picks 18/15 7/6 as the
    best play...
    I'm surprising myself that I'm taking interest in
    a position discussion but it was valuable to see
    that you can't make rocket science out of BG and
    that, not always but often, mine or anyone else's
    judgment is as good as any bg botts.
    MK
    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From Timothy Chow@tchow12000@yahoo.com to rec.games.backgammon on Wed Dec 13 09:13:37 2023
    From Newsgroup: rec.games.backgammon

    On 12/13/2023 2:55 AM, MK wrote:

    Granted they are all very close but shouldn't we
    expect a little more consistency?

    No. Why would anyone expect that when the plays are so close?

    ---
    Tim Chow

    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From MK@murat@compuplus.net to rec.games.backgammon on Fri Dec 22 07:48:18 2023
    From Newsgroup: rec.games.backgammon

    On December 13, 2023 at 7:13:41rC>AM UTC-7, Timothy Chow wrote:
    On 12/13/2023 2:55 AM, MK wrote:
    Granted they are all very close but shouldn't we
    expect a little more consistency?
    No. Why would anyone expect that when the
    plays are so close?
    I'm not complaining about "some inconsistency"
    but I would expect at least "some consistency".
    They are not the same thing. I had said:
    "Except for 3-ply and xg-roller all other plies
    give every different conbination possible."
    Obviously you missed my point that the results
    were "anti-consistent", (as though on purpose).
    Here they are again in more detail:
    1. 2-ply 18/14 eq:+1.0888
    2. 2-ply 18/15 7/6 eq:+1.0865 (-0.0023)
    3. 2-ply 18/16 7/6(2) eq:+1.0834 (-0.0054)
    4. 2-ply 18/17 7/6 7/5 eq:+1.0787 (-0.0101)
    5. 2-ply 18/16 7/5 eq:+1.0754 (-0.0135)
    1. 3-ply 18/14 eq:+1.0841
    2. 3-ply 7/5(2) eq:+1.0839 (-0.0002)
    3. 3-ply 7/6 7/4 eq:+1.0838 (-0.0003)
    4. 3-ply 18/15 7/6 eq:+1.0795 (-0.0045)
    5. 3-ply 18/16 7/6(2) eq:+1.0746 (-0.0095)
    1. XG Roller 18/14 eq:+1.0839
    2. XG Roller 7/6 7/4 eq:+1.0820 (-0.0019)
    3. XG Roller 7/5(2) eq:+1.0808 (-0.0031)
    4. XG Roller 18/15 7/6 eq:+1.0791 (-0.0048)
    5. XG Roller 18/16 7/6(2) eq:+1.0721 (-0.0118)
    1. 4-ply 7/5(2) eq:+1.0803
    2. 4-ply 18/14 eq:+1.0787 (-0.0016)
    3. 4-ply 7/6 7/4 eq:+1.0757 (-0.0045)
    4. 4-ply 18/15 7/6 eq:+1.0742 (-0.0061)
    5. 4-ply 18/16 7/6(2) eq:+1.0682 (-0.0120)
    1. XG Roller+ 18/15 7/6 eq:+1.0766
    2. XG Roller+ 7/5(2) eq:+1.0764 (-0.0003)
    3. XG Roller+ 18/14 eq:+1.0762 (-0.0004)
    4. XG Roller+ 7/6 7/4 eq:+1.0757 (-0.0010)
    5. XG Roller+ 18/16 7/6(2) eq:+1.0676 (-0.0091)
    1. XG Roller++ 18/14 eq:+1.0821
    2. XG Roller++ 18/15 7/6 eq:+1.0818 (-0.0003)
    3. XG Roller++ 7/5(2) eq:+1.0817 (-0.0004)
    4. XG Roller++ 7/6 7/4 eq:+1.0803 (-0.0019)
    5. XG Roller++ 18/16 7/6(2) eq:+1.0754 (-0.0067)
    Forget any consistency among all 5 choices
    but even just the first 2 choices coincide only
    once and quite interestingly betwwn the lowest
    2-ply and the highest XG Roller++ analyses..!
    Contrary to your comments in the other thread,
    they do indeed get more consistent with rollouts
    but again with unexpected exceptions. Let's look:
    First with moves and cube decisions: XG Roller
    1. Rollout-| 18/15 7/6 eq:+1.0844
    2. Rollout-| 7/6 7/4 eq:+1.0841 (-0.0004)
    3. Rollout-| 7/5(2) eq:+1.0827 (-0.0017)
    4. Rollout-| 18/14 eq:+1.0812 (-0.0033)
    5. Rollout-| 18/16 7/6(2) eq:+1.0731 (-0.0114)
    324 Games rolled with Variance Reduction.
    1. Rollout-| 18/15 7/6 eq:+1.0855
    2. Rollout-| 7/6 7/4 eq:+1.0851 (-0.0004)
    3. Rollout-| 7/5(2) eq:+1.0836 (-0.0019)
    4. Rollout-| 18/14 eq:+1.0829 (-0.0026)
    5. Rollout-| 18/16 7/6(2) eq:+1.0745 (-0.0110)
    648 Games rolled with Variance Reduction.
    1. Rollout-| 18/15 7/6 eq:+1.0854
    2. Rollout-| 7/5(2) eq:+1.0849 (-0.0005)
    3. Rollout-| 7/6 7/4 eq:+1.0846 (-0.0007)
    4. Rollout-| 18/14 eq:+1.0837 (-0.0017)
    5. Rollout-| 18/16 7/6(2) eq:+1.0756 (-0.0098)
    1296 Games rolled with Variance Reduction.
    Then with moves and cube decisions: XG Roller+

    1. Rollout-| 7/5(2) eq:+1.0854
    2. Rollout-| 18/15 7/6 eq:+1.0830 (-0.0023)
    3. Rollout-| 18/14 eq:+1.0824 (-0.0030)
    4. Rollout-| 7/6 7/4 eq:+1.0812 (-0.0042)
    5. Rollout-| 18/16 7/6(2) eq:+1.0762 (-0.0092)
    324 Games rolled with Variance Reduction.
    1. Rollout-| 7/5(2) eq:+1.0852
    2. Rollout-| 18/15 7/6 eq:+1.0831 (-0.0021)
    3. Rollout-| 18/14 eq:+1.0814 (-0.0039)
    4. Rollout-| 7/6 7/4 eq:+1.0810 (-0.0042)
    5. Rollout-| 18/16 7/6(2) eq:+1.0764 (-0.0088)
    648 Games rolled with Variance Reduction.
    1. Rollout-| 7/5(2) eq:+1.0846
    2. Rollout-| 18/15 7/6 eq:+1.0834 (-0.0012)
    3. Rollout-| 18/14 eq:+1.0825 (-0.0021)
    4. Rollout-| 7/6 7/4 eq:+1.0817 (-0.0029)
    5. Rollout-| 18/16 7/6(2) eq:+1.0766 (-0.0080)
    1296 Games rolled with Variance Reduction.
    Finally with moves and cube decisions: XG Roller++
    1. Rollout-| 18/15 7/6 eq:+1.0836
    2. Rollout-| 7/5(2) eq:+1.0832 (-0.0004)
    3. Rollout-| 18/14 eq:+1.0826 (-0.0010)
    4. Rollout-| 7/6 7/4 eq:+1.0820 (-0.0016)
    5. Rollout-| 18/16 7/6(2) eq:+1.0763 (-0.0073)
    324 Games rolled with Variance Reduction.
    1. Rollout-| 18/15 7/6 eq:+1.0838
    2. Rollout-| 7/5(2) eq:+1.0837 (-0.0001)
    3. Rollout-| 18/14 eq:+1.0823 (-0.0015)
    4. Rollout-| 7/6 7/4 eq:+1.0819 (-0.0019)
    5. Rollout-| 18/16 7/6(2) eq:+1.0774 (-0.0064)
    648 Games rolled with Variance Reduction.
    1. Rollout-| 7/5(2) eq:+1.0835
    2. Rollout-| 18/15 7/6 eq:+1.0834 (-0.0001)
    3. Rollout-| 18/14 eq:+1.0830 (-0.0004)
    4. Rollout-| 7/6 7/4 eq:+1.0821 (-0.0014)
    5. Rollout-| 18/16 7/6(2) eq:+1.0765 (-0.0070)
    1296 Games rolled with Variance Reduction.
    As would be expected, the results get more precise
    with more trials, (i.e. the equity differences shrink).
    And the different strength bots don't change their
    minds with longer rollouts, except in the last one
    which is the longest rollout for the strongest bot..!
    (It took over 12 hours on low priority setting but it
    was well worth it as we can see :)
    It all looks like I am vindicated that ER/PR would
    be horse muffins even if calculated by doing long
    enough strong rollouts but they are worthless cow
    pies even more based on simpler analyses done
    during the games even by XG Roller++ :( So, just
    go with 2-ply and be happy :))
    MK
    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From Timothy Chow@tchow12000@yahoo.com to rec.games.backgammon on Sat Dec 23 08:34:29 2023
    From Newsgroup: rec.games.backgammon

    On 12/22/2023 10:48 AM, MK wrote:

    Contrary to your comments in the other thread,
    they do indeed get more consistent with rollouts
    but again with unexpected exceptions. Let's look:

    There is nothing unexpected about your results. You have
    omitted the most important quantity from your rollout results,
    which is the confidence interval. It is to be expected that
    there will be random fluctuations within, say, two times the
    width of the confidence interval.

    ---
    Tim Chow

    --- Synchronet 3.21b-Linux NewsLink 1.2
  • From MK@murat@compuplus.net to rec.games.backgammon on Tue Dec 26 12:23:02 2023
    From Newsgroup: rec.games.backgammon

    On December 23, 2023 at 6:35:24rC>AM UTC-7, Timothy Chow wrote:
    On 12/22/2023 10:48 AM, MK wrote:

    Contrary to your comments in the other thread,
    they do indeed get more consistent with rollouts
    but again with unexpected exceptions. Let's look:
    There is nothing unexpected about your results.
    You have omitted the most important quantity
    from your rollout results, which is the confidence
    interval. It is to be expected that there will be
    random fluctuations within, say, two times the
    width of the confidence interval.
    Confidence interval is a result, not a parameter of
    the rollouts I am doing to compare their results. If
    I use it as a paramater, it will defeat my purpose.
    But there may be something I don't understand in
    what you said and would like to hear you explain.
    After achieving my own challenge of consistently
    beating the bots behind me, I gradually took off my
    "brain mask" for the past few years, being no longer
    worried about keeping my mind uninfected by what
    I call "fancyful bullshits" like equities, cube points,
    etc. I feel fairly safe by now that I won't become one
    of you all who try to play like the bots, (thus not be
    able to beat them at their own game), and I'm more
    willing to understand those "fancyful bullshits" in
    order to better argue against them.
    I have always argued that cube skill formulas to
    extrapolate cubeful equities and match tables to
    extrapolate "matchful" equities injected systematic
    errors of unknown magnitute, each by itself or both
    in combination.
    I have also argued that they would accumulate in
    proportion to the lengths of rollouts.
    In my above rollouts, I was trying to isolate random
    and systematic errors by comparing many rollouts
    of different lengths and/or different plies. I realize
    that a few rollouts aren't significant but it's a start.
    Perhaps you can help me out with it by suggesting
    a better way doing and comparing, cubeless/cubeful
    and matchless/matchful (single games/multi-game
    matches).
    MK
    --- Synchronet 3.21b-Linux NewsLink 1.2