1. Forum
  2. Usenet
  3. REC.GAMES.BACKGAMMON

  • Awed by the bots: "54 safe or bold?" expanded edition

    From MK@21:1/5 to All on Thu Mar 23 22:11:09 2023
    This article is what I was about to post when I had made the
    bad choice of responding instead to Axel's positions, enticed
    by his misleading "hints" about "score".

    In one of my previous posts, I had written:

    https://groups.google.com/g/rec.games.backgammon/c/lcOpsz6CJgk/m/Ad5tFc0lBAAJ

    On March 6, 2023 at 3:29:23 AM UTC-7, MK wrote:

    XGID=-ABbBCC-B----a---bbbbBbb--:0:0:1:54:0:0:3:0:10
    X to play 54

    I spent over half hour trying all kinds of combinations of
    match lengths and match scores and cube values to see
    if Ex-Gee would make a move other than 6/1 5/1.

    In all that time, I came accross only a few match lengths
    and match scores and cube values combinations that
    Ex-Gee actually played 21/16 5/1.

    As you may have guesses, I was clueless as to what caused
    Ex-Gee to play anything other than 6/1 5/1 in those situations.

    For now, I won't share those situations with you here in order
    to see if anyone of you can come up with similar examples
    and/or if any of the AI-bot developers here can directly give
    any explanations based on their having programmed some
    algorithms into their gamblegammon bots.


    Nobody posted any examples or formulas since then but
    Axel's explaining his "hints" (as not about score because
    his examples didn't have "lop-sided scores" and the cube
    values weren't "high"), indicates that he must have at least
    experimented with different scores/cubes even if he hasn't
    posted about them.

    Indeed, in the subject position above, the bots make "bold"
    moves only if the score is very lop-sided and if at the same
    time the cube is high enough but not too high (kind of in a
    "Goldilocks Zone").


    Before starting to talk about matches, let me reiterate that
    the "54 safe or bold?" was a stupid question to ask for the
    money game example position and sure enough both Ex-Gee
    and Noo-BG always play 6/1 5/1 regardless of with/without
    jacoby, cube value, cube ownership and equity.

    Since I'm clueless about estimating equities, cubing points,
    MET's, etc. I had to experiment by poking around. I tried score
    and cube combinations in matches from 7 to 33 points, but
    focused most on 15 and 25 matches. Here are my generalized
    observations with some actual examples, following which I will
    make a few final comments.


    All examples with both bots using Rockwell/Kazaross MET
    but Noo-BG played the same positions somewhat differently.

    (Notice that Noo-BG allows centered cube values > 1 and it
    analyzes them as though the player to move owns the cube.)


    With the cube at 1, 2, 4, 32, 64, ... owned by either side, both
    Noo-BG and Ex-Gee always played 6/1 5/1 at any score and
    at any match length.

    With the cube at 8, owned by either side, both Noo-BG and
    Ex-Gee played 6/1 5/1 or 21/16 5/1 or 21/12 or 8/4 6/1
    depending on score and match length combinations.

    With the cube at 16, owned by either side, Ex-Gee played
    6/1 5/1 or 21/16 5/1 or 21/12 or 8/4 6/1 depending on the
    score and match length combinations, but Noo-BG always
    played 6/1 5/1.


    Here are some Ex-Gee examples with cube at 16:

    XGID=-ABbBCC-B----a---bbbbBbb--:4:1:1:54:2:22:0:25:10
    MWC: 24.122% P: 51.3 6.1 0.1 - O: 48.7 8.5 0.2
    At score X=2, O=22 it plays 6/1 5/1

    XGID=-ABbBCC-B----a---bbbbBbb--:4:1:1:54:1:22:0:25:10
    MWC: 20.478% P: 42.1 6.9 0.1 - O: 57.9 9.4 0.3
    At score X=1, O=22 it plays 8/4 6/1

    XGID=-ABbBCC-B----a---bbbbBbb--:4:1:1:54:0:22:0:25:10
    MWC: 17.720% P: 46.9 6.7 0.1 - O: 53.1 11.1 0.4
    At score X=0, O=22 it plays 21/12

    XGID=-ABbBCC-B----a---bbbbBbb--:4:1:1:54:0:21:0:25:10
    MWC: 21.949% P: 42.0 7.0 0.1 - O: 58.0 9.4 0.3
    At score X=0, O=21 it plays 8/4 6/1


    Here are some Ex-Gee examples with cube at 8:

    XGID=-ABbBCC-B----a---bbbbBbb--:3:1:1:54:2:22:0:25:10
    MWC: 4.870% P: 48.9 6.3 0.1 - O: 51.1 10.5 0.4
    At score X=2, O=22 it plays 21/12

    XGID=-ABbBCC-B----a---bbbbBbb--:3:1:1:54:1:22:0:25:10
    MWC: 3.983% P: 50.6 6.3 0.1 - O: 49.4 9.8 0.4
    At score X=1, O=22 it plays 21/12

    XGID=-ABbBCC-B----a---bbbbBbb--:3:1:1:54:0:22:0:25:10
    MWC: 3.224% P: 51.0 6.2 0.1 - O: 49.0 9.7 0.4
    At score X=0, O=22 it plays 21/12

    XGID=-ABbBCC-B----a---bbbbBbb--:3:1:1:54:0:21:0:25:10
    MWC: 4.733% P: 42.1 6.9 0.1 - O: 57.9 9.3 0.3
    At score X=0, O=21 it plays 8/4 6/1


    Here are some Noo-BG examples with cube at 16:

    GNUbg ID: NtuGAAzN7gwABg:VIkyA2ABEAAA
    MWC: 25.08%
    At score X=2, O=22 it plays 8/4 6/1

    GNUbg ID: NtuGAAzN7gwABg:VIkyA2ABCAAA
    MWC: 22.01%
    At score X=1, O=22 it plays 8/4 6/1

    GNUbg ID: NtuGAAzN7gwABg:VIkyA2ABAAAA
    MWC: 19.35%
    At score X=0, O=22 it plays 21/12

    GNUbg ID: NtuGAAzN7gwABg:VIkyA1ABAAAA
    MWC: 23.28%
    At score X=0, O=21 it plays 8/4 6/1


    Here are some Noo-BG examples with cube at 8:

    GNUbg ID: NtuGAAzN7gwABg:U4kyA2ABEAAA
    MWC: 5.33%
    At score X=2, O=22 it plays 21/16 5/1

    GNUbg ID: NtuGAAzN7gwABg:U4kyA2ABCAAA
    MWC: 4.34%
    At score X=1, O=22 it plays 21/16 5/1

    GNUbg ID: NtuGAAzN7gwABg:U4kyA2ABAAAA
    MWC: 3.54%
    At score X=0, O=22 it plays 21/16 5/1

    GNUbg ID: NtuGAAzN7gwABg:U4kyA1ABAAAA
    MWC: 5.12%
    At score X=0, O=21 it plays 21/16 5/1


    Obviously the bots use the same/similar fancifully elaborate
    formulas to calculate the MWC's using the MET's, in order to
    precisely pick the best moves.

    How could I not be totally impressed by the above garbage
    produced by Ex-Gee and Noo-BG...? :)

    What I can't figure out though is whether the scores need to
    be so highly lop-sided in addition to and in relation to high
    "Goldilocks cube values", or the formulaes themselves are
    lop-sided...?? :(

    MK

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)