• Re: Tesla backgammon?

    From WF Frantz@21:1/5 to Tim Chow on Fri Jun 2 14:23:52 2023
    On Saturday, May 23, 2020 at 9:17:16 AM UTC-7, Tim Chow wrote:
    I bought Robertie's new book and so got onto his mailing list.
    One thing he mentioned recently is that the Tesla backgammon
    program has apparently improved from being a terrible player
    to being a challenging player. Does anyone here know anything
    about this?

    From a business point of view, it would seem much better to
    have a weak program than a strong program, since you want to
    make the customer feel smart.

    ---
    Tim Chow
    Yes. The game maker claims that he used AI to improve, but that is total bull. It now cheats like crazy.

    Even the 2023 version is rigged, but not how you think. After nearly 10,000 dice rolls, this is what I found when keeping track of rolls and getting back onto the board after being captured (clearing the bar):

    The designer claims that the new game was put into learning mode and got a lot better, which is a total lie. He never did that. It got better by improving how it hides cheating.

    Doubles are now OK. But ...

    Getting out of the Bar (pieces taken) is rigged in favor of the computer. The computer beats the expected odds of getting back onto the board while the player is more likely than expected to not get back onto the board.

    It is more subtle about the dice. It equally rolls more 4s for the computer and more 1s for the player. In Super Tesla mode, the dice rolls are now about 13% higher for the computer, so the computer has higher odds of completing first. This is subtle.
    The average roll should be 7. But the computer average is 7.2 and player 6.8. And by increasing the number of 4s for the computer and 1s for the player, the 1s are a definite handicap because blockades are normally face to face making the use of 1s
    limited.

    The game is still stupid on the end game, which is the only reason why players continue to beat the computer in Super Tesla mode about half the time. The game has little situational awareness as to: (a) minimizing a 2 or 3 point penalty for losing, (b)
    focusing on getting pieces home once all pieces have passed each other, and (c) prioritizing clearing home and balancing home when a piece can't be cleared.

    I wish that Tesla would not reward game makers that use cheating especially when they claim that they used AI. I wish Tesla would reward game makers with a work ethic that represents the work ethic of Tesla.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From pepstein5@gmail.com@21:1/5 to Tim Chow on Sat Jun 3 01:42:26 2023
    On Saturday, May 30, 2020 at 5:20:43 PM UTC+1, Tim Chow wrote:
    On Saturday, May 30, 2020 at 12:17:25 PM UTC-4, I wrote:
    Yes, I was rather surprised when I first learned that 20 million games was enough training data for a decent bot
    Just checked Tesauro's paper and found that TD-Gammon Version 2.1 used only 1.5 million games. That wouldn't be considered a strong bot today but it still illustrates the point that 20 million games is a reasonable amount of data.

    If I remember rightly, it was an open question whether TD-Gammon was better/worse/equal
    when compared to the world's strongest humans.
    Of course, with the huge luck factor in backgammon, it can be very hard to ascertain these things.
    I think that now, there's a strong consensus that both XG and Gnu are much stronger than any human,
    and I'm sure there are plenty of other bots that surpass human play.

    Paul

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From pepstein5@gmail.com@21:1/5 to Tim Chow on Sat Jun 3 01:38:43 2023
    On Saturday, May 30, 2020 at 5:17:25 PM UTC+1, Tim Chow wrote:
    On Saturday, May 30, 2020 at 9:20:32 AM UTC-4, bgbl...@googlemail.com wrote:
    20 Million games is enough for a quite decent player......
    if you know what you do. Maybe they should ask one who knows what to do ;)
    Yes, I was rather surprised when I first learned that 20 million games was enough training data for a decent bot, because I was naively guessing that it would require ten times as much as that. But apparently, even AlphaZero chess used only 44 million chess games, so these neural nets are somehow able to generalize incredibly well from (what seems to me like) relatively little data.
    But as Tim pointed out: for the average customer to keep happy playing strength might not the right way to go. Does one remember the "Big Bang Board Games" Apple put on their Macs a decade or so ago. Much Bling-Bling and abysmal game play....
    I have posted here before about the terrible level of play of the backgammon game on JetBlue. But Stick has a point...if the computer player is *too* weak then that can also be unappealing.

    FWIW, my intuition was also that the low tens of millions would be far too few for
    complex games like chess or backgammon.

    You can take the man out of combinatorial game theory, but you can't take combinatorial
    game theory out of the man.

    Paul

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Frank Berger@21:1/5 to peps...@gmail.com on Sat Jun 3 13:28:32 2023
    peps...@gmail.com schrieb am Samstag, 3. Juni 2023 um 10:38:45 UTC+2:

    FWIW, my intuition was also that the low tens of millions would be far too few for
    complex games like chess or backgammon.

    How many games a human needs ;) When you play with nets that are trained ony a couple of thousand games, you can watch how they graps more and more concepts.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Frank Berger@21:1/5 to peps...@gmail.com on Sat Jun 3 13:35:31 2023
    peps...@gmail.com schrieb am Samstag, 3. Juni 2023 um 10:42:28 UTC+2:

    I think that now, there's a strong consensus that both XG and Gnu are much stronger than any human,
    and I'm sure there are plenty of other bots that surpass human play.

    How many are plenty?
    As in an earlier post:
    - XG
    - GnuBG
    - BGBlitz
    - Snowie
    are very very probably better than any human on average

    TD-Gammon/Jellyfish/Motif? Are you sure that these are better than *any* human. If Mochy, ZZ and Ueda play together as a team against them, on whom you would bet? I'm *very* unsure

    Do I miss some other?

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Frank Berger@21:1/5 to All on Sat Jun 3 13:36:49 2023
    WF Frantz schrieb am Freitag, 2. Juni 2023 um 23:23:53 UTC+2:
    would you mind to make your data available? cheating with the average would be a more smart way to do it

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From pepstein5@gmail.com@21:1/5 to Frank Berger on Sun Jun 4 02:21:47 2023
    On Saturday, June 3, 2023 at 9:35:32 PM UTC+1, Frank Berger wrote:
    peps...@gmail.com schrieb am Samstag, 3. Juni 2023 um 10:42:28 UTC+2:

    I think that now, there's a strong consensus that both XG and Gnu are much stronger than any human,
    and I'm sure there are plenty of other bots that surpass human play.
    How many are plenty?
    As in an earlier post:
    - XG
    - GnuBG
    - BGBlitz
    - Snowie
    are very very probably better than any human on average

    TD-Gammon/Jellyfish/Motif? Are you sure that these are better than *any* human. If Mochy, ZZ and Ueda play together as a team against them, on whom you would bet? I'm *very* unsure

    Do I miss some other?

    Motif was terrible last time I played it. It may have improved since then.
    I don't know these issues particularly well. I suppose "sure" was probably the wrong word.
    The team hypothesis does not address the question because "better than any human" means
    "better than any individual human".
    It's plausible (at least to me) that a team of the 20 best humans could play better than XG.

    Paul

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From pepstein5@gmail.com@21:1/5 to peps...@gmail.com on Sun Jun 4 03:55:43 2023
    On Sunday, June 4, 2023 at 10:21:48 AM UTC+1, peps...@gmail.com wrote:
    ....
    Motif was terrible last time I played it. It may have improved since then.
    ...

    I wanted to test the latest Motif but I was stuck on the Java requirements.
    I got this message: "If this message appears, then your web browser does not run Java and you will not be able to play Motif backgammon."

    I'm sure this has absolutely nothing to do with Motif because other java-requiring sites exhibit exactly the same issue.
    I'd like to get this working on Google Chrome but I couldn't get it working on Microsoft Edge or Firefox either.
    I've tried googling instruction sets without success.
    Most of the instructions appear to assume menu functions which don't seem to exist.
    I am using Windows 64 x64

    Many thanks for your help.

    I do have some partial successes as follows:
    Successful installation indicated as below:
    java -version
    java version "1.8.0_371"
    Java(TM) SE Runtime Environment (build 1.8.0_371-b11)
    Java HotSpot(TM) Client VM (build 25.371-b11, mixed mode)

    Paul

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Frank Berger@21:1/5 to peps on Sun Jun 4 14:07:51 2023
    peps schrieb am Sonntag, 4. Juni 2023 um 12:55:44 UTC+2:
    I wanted to test the latest Motif but I was stuck on the Java requirements. I got this message: "If this message appears, then your web browser does not run Java and you will not be able to play Motif backgammon."
    Java in the browser isn't supported in Browsers nowadays (for good reasons but this is a long long story)

    I checked when Java was thrown out of java itself, but it was in Java 9 so your Java 1.8 should work. The program would be the "appletviewer"... but it don't
    so before investing a lot of time I thought: let's try it with an old environment, but on my recent boxes no browser supports the Java-plugin and on the older (OS/2 in a VM, iMac from 2000) they couldn't handle the new algos for https/ssl :(

    finally I checked it with curl whether the files are still there.... and they are. I downloaded them and packed them together at.
    https://bgblitz.com/download/misc/Motif.zip

    If you
    - download it,
    - unpack it at any convenient place
    - cd the commandline in this directory
    - and call
    appletviewer ./motif.html

    it should work (at least on a Linux and a Mac it works. To lazy to start WIndows).

    BTW there is a webpage , where Hans-Jürgen Schäfer collects elo numbers from every BG bot he can get:
    https://mustrum.de/blot.html

    It is in German, but if you scroll down you'll find the elo numbers. I don't know his exact procedure, but the number looks plausible to me. The most bias is probably due to setting max elo to 2050.
    Astonishing/amazing how bad some bots are....

    best
    Frank

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From pepstein5@gmail.com@21:1/5 to Frank Berger on Mon Jun 5 00:43:43 2023
    On Sunday, June 4, 2023 at 10:07:53 PM UTC+1, Frank Berger wrote:
    peps schrieb am Sonntag, 4. Juni 2023 um 12:55:44 UTC+2:
    I wanted to test the latest Motif but I was stuck on the Java requirements.
    I got this message: "If this message appears, then your web browser does not run Java and you will not be able to play Motif backgammon."
    Java in the browser isn't supported in Browsers nowadays (for good reasons but this is a long long story)

    I checked when Java was thrown out of java itself, but it was in Java 9 so your Java 1.8 should work. The program would be the "appletviewer"... but it don't
    so before investing a lot of time I thought: let's try it with an old environment, but on my recent boxes no browser supports the Java-plugin and on the older (OS/2 in a VM, iMac from 2000) they couldn't handle the new algos for https/ssl :(

    finally I checked it with curl whether the files are still there.... and they are. I downloaded them and packed them together at.
    https://bgblitz.com/download/misc/Motif.zip

    If you
    - download it,
    - unpack it at any convenient place
    - cd the commandline in this directory
    - and call
    appletviewer ./motif.html

    it should work (at least on a Linux and a Mac it works. To lazy to start WIndows).

    BTW there is a webpage , where Hans-Jürgen Schäfer collects elo numbers from every BG bot he can get:
    https://mustrum.de/blot.html

    It is in German, but if you scroll down you'll find the elo numbers. I don't know his exact procedure, but the number looks plausible to me. The most bias is probably due to setting max elo to 2050.
    Astonishing/amazing how bad some bots are....

    best
    Frank

    Thanks a lot for your help. I'll let you know how I get on.

    Paul

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From MK@21:1/5 to peps...@gmail.com on Mon Jun 5 01:53:50 2023
    On June 3, 2023 at 2:42:28 AM UTC-6, peps...@gmail.com wrote:

    On May 30, 2020 at 5:20:43 PM UTC+1, Tim Chow wrote:

    On May 30, 2020 at 12:17:25 PM UTC-4, I wrote:

    Yes, I was rather surprised when I first learned
    that 20 million games was enough training data
    for a decent bot

    What one deems "decent" is relative and also would
    depend on being a decent bot at what kind of games.
    Those were all cubeless one-pointers.

    Just checked Tesauro's paper and found that
    TD-Gammon Version 2.1 used only 1.5 million
    games. That wouldn't be considered a strong
    bot today but it still illustrates the point that 20
    million games is a reasonable amount of data.

    It wasn't a strong bot back then either. Not 1.5 nor
    even 20 million "cubeless one-pointer games" can
    be a "reasonable amount of data" for for cubeful
    money and/or match play, unless some jackoffski
    formulas, MET's, etc. are substituted for the lack
    of training and unless "the small incestious circle
    of gambling gamblegammon players" subscribed
    to it.

    If I remember rightly, it was an open question
    whether TD-Gammon was better/worse/equal
    when compared to the world's strongest humans.

    It was only compared to a few humans considered
    strongest among a handful of gamblers who rated
    themselves against one another but at least a real
    comparison was done by those humans actually
    playing against the TD-Gammon.

    I think that now, there's a strong consensus that
    both XG and Gnu are much stronger than any human

    Consensus doesn't mean based on factual data. The
    question is still open and even more so now than back
    then because unlike with TD-Gammon, there has been
    no serious comparing to humans through actual play
    against those bots with incentives enough for humans
    to really try to beat them.

    Just the consensus among billions of insect brained
    people that gods, angels, spirits, etc. exist doesn't
    make their fantasies into facts...

    MK

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From MK@21:1/5 to Frank Berger on Mon Jun 5 02:29:53 2023
    On June 3, 2023 at 2:35:32 PM UTC-6, Frank Berger wrote:

    How many are plenty?
    As in an earlier post:
    - XG
    - GnuBG
    - BGBlitz
    - Snowie

    Why do you keep injecting your garbage
    bot BG-bzzt into the list at evert occasion?

    Even though it's dated by now, Snowie still
    belongs there because it's the only bot that
    disagrees with Ex-Gee and Noo-BG at times.

    Nobody other than yourself does hardly ever
    mention, let alone praise, your garbage bot
    BG-bzzt. :(

    Add some unique features to it to make it a
    little more useful and valuable, such as the
    random rollouts that I've been suggesting for
    some years now.

    In BGO, they are talking about PC's that can
    do 46 million positions/second in Ex-Gee.

    The 1,296,000 trials random rollout that I did
    on my 3.8 million positions/second PC had
    taken less than 5 hours. That means with a
    Ryzen 9 7950X CPU you can do a ten times
    longer random rollout, i.e. 12,960,000 trials,
    in just 4 hours.

    If a 12 million 960 thousand trials of random
    cubeful rollout cames up with different best
    plays than a 1,296 trials Ex-Gee rollout with
    4-ply checker and XGroller+ cube settings,
    will you all not accept it as more credible..??

    What are you all afraid of to not add a simple
    feature to your bots that could demonstrate
    them to be pieces of shit in a matter of a few
    dozens of random play rollouts that would at
    most take a few days...?

    You already know it! You can delay accepting
    it openly for now but you can't prevent it forever.
    It's only a matter of time...

    MK

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From Frank Berger@21:1/5 to All on Mon Jun 5 03:49:51 2023
    MK schrieb am Montag, 5. Juni 2023 um 11:29:54 UTC+2:

    Nobody other than yourself does hardly ever
    mention, let alone praise, your garbage bot
    BG-bzzt. :(
    Not that I believe that you were reachable for arguments or facts, but feel free to make some matches with BGBlitz (I suggest 4- or 5-ply) or BGBlitz against itself and feed them to XG. You don't regard PR as reasonable IIRC but Snowie is a bit over 1
    and BGBlitz at 5-ply is around 0.6 and Mochy is about 2.3.

    Or you may remember the Othello 2023 Quiz posted a few days ago? What you call "garbage" solves
    - 8 out of 10 with 1-ply
    - 9 out of 10 wth 2-ply
    - 10 out of 10 with 5ply (the 1st position is really a beauty)
    I'm confident that you wont regard that as showing anything, but others might find that interesting. BTW one doesn't need a license to check that, just drag&drop the XGID on the board.

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)
  • From MK@21:1/5 to Frank Berger on Tue Jun 6 01:58:16 2023
    On June 5, 2023 at 4:49:53 AM UTC-6, Frank Berger wrote:

    MK schrieb am Montag, 5. Juni 2023 um 11:29:54 UTC+2:

    Nobody other than yourself does hardly ever
    mention, let alone praise, your garbage bot
    BG-bzzt. :(

    ... feel free to make some matches with BGBlitz
    (I suggest 4- or 5-ply) or BGBlitz against itself and
    feed them to XG. You don't regard PR as reasonable
    IIRC but Snowie is a bit over 1 and BGBlitz at 5-ply
    is around 0.6 and Mochy is about 2.3.

    My point is that your bot isn't popular. I don't care
    if it's almost as strong as Ex-Gee. Even yourself
    measure your bot against Ex-Gee, Noo-BG, Snowie.
    They're all the same. Your bot doesn't offer anything
    significantly different than other bots.

    You seem to be making quite an effort to improve
    and keep your bot current but you won't even take
    a small suggestion that could help it break away
    from the inbread dog pack. :( At time, I try to help
    you out with some ideas to no avail...

    Or you may remember the Othello 2023 Quiz posted
    a few days ago? What you call "garbage" solves
    - 8 out of 10 with 1-ply
    - 9 out of 10 wth 2-ply
    - 10 out of 10 with 5ply (the 1st position is really a beauty)

    Big fucking deal. They were all rolled out with Ex-Gee
    "1296 Games rolled with Variance Reduction.
    "Moves: 3-ply, cube decisions: XG Roller

    As if I had nothing better to do, I rolled them all using:
    "1296 Games rolled with Variance Reduction.
    "Moves and cube decisions: 1-ply

    Guess what? Ex-Gee 1-ply got 10 out of 10 right.

    (BTW: this validates Tim's claims that Othello pics
    positions that won't vary based on bot and/or rollot
    settings).

    Furthermore, if the quiz was done using BG-bzzt, it's
    very likely that your 1-ply would get them all right.

    You people just can't understand the absurdity of it
    all... :(

    But more importantly, why are you ignoring/avoiding
    my suggestion to you?

    I just looked at your bot's rollout settings. I only see
    1, 2, 3-ply, no 5-ply. But that's not important. Random
    rollouts can make your bot stand out from the rest.
    How can't you see that? And when it's so easy to do.

    All you need to do is modify one line and add 5 lines
    to your code. Here are the steps:

    1- Modify the line that sets the size of your drop down
    list array from 3 to 4.

    2- Add a new line to add a new rollout level, (i.e. random),
    to the array.

    3- I assume you have a block of code, perhaps with calls
    to subroutines, which decides if the player can/should
    initiate a cube action and if/how he must respond to a
    cube action. Just before that block add a line to execute
    it contitionally, (i.e. IF ROLLOUT-TYPE NOT = "RANDOM")
    and right after it add a line to make a random cube
    decision, (i.e. ELSE CUBE-ACTION = RAND(0 or 1)). That's
    it. Just let the rest of the cube logic execute as normal.

    4- To decide how to move, I assume your code first fills
    an array with all possible moves and then evaluates and
    ranks them to pick the best move. Right after filling the
    array, again add a line to execute that block of code
    conditionally, (i.e. IF ROLLOUT-TYPE NOT = "RANDOM")
    and just before the code that executes the chosen play,
    add a line to pick a random move from the array, (i.e.
    ELSE BEST-MOVE = ARRAY-ELEMENT(RAND(1 to ARRAY
    SIZE). That's it. Just let the rest of the checker play logic
    execute as normal.

    Okay, my sample lines to add are very sloppy but I'm
    sure you get the idea. Just modify 1 line and add 5
    new lines. How long would this take for you to do?
    20 minutes? 30 minutes? 1 hour?

    What's your excuse for refusing to add this simple
    yet potentially jackoffski cube skill debunking,
    gamblegammon shaking feature to your garbage
    bot...??

    Ex-Gee is obviously an abandoned product by now
    but I don't understand why Noo-BG team doesn't add
    this feature to their bot, which would be just as easy
    to do as adding for you to add it to your bot.

    Remember that you all won't be able to prevent the
    unevitable. Save yourselves the agony and the pain.
    Bite the bullet. Get done and get over with it. Sooner
    the better for you all. Trust my friendly advice... ;)

    MK

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