| |
BGonline.org Forums
Murat's HypestGammon - Whatever happened to Zare, ZBot, Chow..?
Posted By: MK In Response To: Whatever happened to Zare, ZBot, Chow..? (Timothy Chow)
Date: Tuesday, 7 July 2026, at 7:25 a.m.
>> Why don't you invite your "colleague" here to start up a thread about it?
> There's nothing much to discuss until he has concrete results to share.
I meant discussing how to approach training a bot through cubeful self-play.
> I will say, though, that because he was interested in the doubling cube, he did one preliminary experiment, with an extremely simplified version of backgammon where there was very little to decide other than if and when to double.
For the ones who don't know the history of this subject, he is talking about "Murat's HypestGammon" that I invented in 2020 to isolate the so-called "cube skill".
It's the simplest gamblegammon variant played with one checher each and rolling a single die. Since all moves are forced, it requires 0% checker skill. Thus any remaining skill is 100% "cube skill".
Tim had claimed that he could calculate cubeful equities in HypestGammon exactly. And I had dared him and "his colleague" to prove his calculations would be accurate through empirical data (i.e. from cubeful self-play). Apparently they have not being able to do that yet after over six years.
Before we go any further on this subject, I urge all of you to read (or re-read to refresh your mind if you had read it in the past) at least this one thread from RBG:
https://groups.google.com/g/rec.games.backgammon/c/h03ubUZEd2s/m/wNOvV2HzBQAJ
Try to keep an open mind and read the entire thread. I can't promise but you may learn something from it and surprise yourselves.
> He found that if he coded up the problem naïvely, AlphaZero was not able to learn good cube handling. It would double randomly, which meant doubling a lot. That meant that the scores would vary wildly from one game to the next, creating so much "noise" that it drowned out the "signal." In order to get any meaningful results, he had to tweak some parameters, e.g., I think he tried setting the initial default probability of doubling to some low number, to discourage it from doubling too aggressively.
These are the things that would be worth discussing. I'm sure there are people here who are as knowledgeable in AI bot training as your colleague and they all may come up with solutions collectively, to whatever problems you are talking about using the words "naïvely", "noise", "signal", "meaningful results", "tweak some parameters", "initial default probability of doubling", etc.
My argument has been all long that the so-called "cube skill" based on cubeful equities extrapolated using jackoffski formulas was mostly fantasy (i.e. in early and middle stages of games), with only a fraction of it that can be called real skill (i.e. in the late stages of games).
I have demonstrated this with my mutant-bot experiments using my own "fartoffski" (purposefully arbitrary and primitive) cube formulas based on game stages that performed nearly as well as you folk's "jackoffski" formulas. See the "Deforming the cube skill theory dogmatism" section of my backgammon site:
https://www.montanaonline.net/backgammon/index.php
I have long argued and I am still arguing that an unbiased cubefully trained AI will prove me right.
It seems like what your colleague is trying to do is to manipulate the training to produce predesired "meaningful results", which is obviously futile. Why not let the chips fall where they may and live with the reality, instead of trying to perpetuate fantasies..?
> There is, at minimum, a lot of jiggling around of hyperparameters and training algorithms that goes into developing a world-class bot.
That is because you all can't free yourselves from thinking that "a world-class bot" needs to play against itself like XG and GnuBG play against themselves. What if a bot plays in a way that doesn't seem "meaningful" to you but achieves better "actual results" against other bots and/or humans..?
> I'm sure there were many, many iterations of "try this, and if the bot plays like crap, change something and try again."
If a chess or backgammon bot can be better than any human, how can an inferior human tell if a superior bot plays like crap? Before anything else, you need to understand this question first...
MK
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.