If you wanted to analyse a bunch of games to
accumulate stats on certain types of checker
plays or cube actions at various stages of the
game, (i.e. opening, early, middle, late, closing),
how would you try to recognise those stages?

Based on the average length of a large number
of games of all types, (let's say 25-30 rolls per
player), the first 5-10 rolls may be considered
the opening and early stages but after that it's
not all that easy, especially in gamblegammon
where games can end short abrubtly.

However, since we are talking about averages
over large enough number of samples, would it
still be good enough to say that? (For example,
after 25 rolls we can assume that we are in the
bearing off stage..?)

How else better can this be done? Can we tell
anything from the psition ID's? Any other ideas?

MK

MK <murat@compuplus.net> writes:

> If you wanted to analyse a bunch of games to accumulate stats on
> certain types of checker plays or cube actions at various stages of
> the game, (i.e. opening, early, middle, late, closing), how would you
> try to recognise those stages?

GNU Backgammon has several evaluators, IIRC at least one for pure races
and one for crashed positions apart from the "standard" evaluation.

> after 25 rolls we can assume that we are in the bearing off stage..?

No. It might be a backgame or a last man hanging out for a shot etc.

> Can we tell anything from the psition ID's?

Obviously, because it encodes the full state of checkers on the board.

Best regards

Axel

On 12/14/2022 7:40 AM, Axel Reichert wrote:
> MK <murat@compuplus.net> writes:
>
>> If you wanted to analyse a bunch of games to accumulate stats on
>> certain types of checker plays or cube actions at various stages of
>> the game, (i.e. opening, early, middle, late, closing), how would you
>> try to recognise those stages?
>
> GNU Backgammon has several evaluators, IIRC at least one for pure races
> and one for crashed positions apart from the "standard" evaluation.
>
>> after 25 rolls we can assume that we are in the bearing off stage..?
>
> No. It might be a backgame or a last man hanging out for a shot etc.
>
>> Can we tell anything from the psition ID's?
>
> Obviously, because it encodes the full state of checkers on the board.

Well, it's not possible to tell from the state of the checkers on the
board how many moves have occurred since the start of the game. In
theory, one could infer the *minimum* number of moves that have occurred
since the start of the game, but ascertaining the minimum number is a
potentially difficult computational problem.

What I think Murat is asking is how to quickly process the transcript
of a large number of games and produce several separate lists of
positions, categorized according to type. He suggested "opening,
early, middle, late, closing" but you could imagine wanting other
categories, such as noncontact positions, bearoff against contact,
backgames, etc.

Depending on how fine a classification one wants and how accurately
one wants the classification to be, this could be a fairly challenging
programming task. Recognizing non-contact positions is fairly easy.
(Though even the definition of "contact" can be a little subtle, as
this article by Tom Keith

https://bkgm.com/articles/Keith/Hyper09/index.html

demonstrates.) If you have the game transcript, then the opening is
pretty easy too; you just have to decide when to cut off the opening
(first four rolls of the game, perhaps?). But other types of positions
aren't necessarily so easy to codify into a set of simple rules that
you can program easily.

---
Tim Chow

On December 14, 2022 at 5:40:11 AM UTC-7, Axel Reichert wrote:

> MK <mu...@compuplus.net> writes:

>> If you wanted to analyse a bunch of games to
>> accumulate stats on certain types of checker
>> plays or cube actions at various stages of the
>> game, (i.e. opening, early, middle, late, closing),
>> how would you try to recognise those stages?

> GNU Backgammon has several evaluators, IIRC
> at least one for pure races and one for crashed
> positions apart from the "standard" evaluation.

I don't think those will help me since what I had
in mind is more looking back than forward.

>> after 25 rolls we can assume that we are in the
>> bearing off stage..?

> No. It might be a backgame or a last man hanging
> out for a shot etc.

As a prelude to my question, I had said "talking
about averages over large enough number of
samples"...

>> Can we tell anything from the psition ID's?

> Obviously, because it encodes the full state of
> checkers on the board.

I'm not sure "state of checkers on the board"
will be enough. Can you elaborate how?

MK

On December 14, 2022 at 6:48:44 AM UTC-7, Tim Chow wrote:

> On 12/14/2022 7:40 AM, Axel Reichert wrote:

>> MK <mu...@compuplus.net> writes:

>>> Can we tell anything from the psition ID's?

>> Obviously, because it encodes the full state
>> of checkers on the board.

> What I think Murat is asking is how to quickly

"Quickly" is the key word that I should have used
to be clearer. Thanks for filling it in.

> process the transcript of a large number of
> games and produce several separate lists of
> positions, categorized according to type. He
> suggested "opening, early, middle, late, closing"

I was thinking of something like plotting cube
decisions and cube values through the length
of games but nothing more complicated than
that at least initially for now.

> ..... you just have to decide when to cut off the
> opening (first four rolls of the game, perhaps?).
> But other types of positions aren't necessarily
> so easy to codify into a set of simple rules that
> you can program easily.

I was about to suggest another, (more advanced
but still simple to do), mutant bot experiment for
Axel on cube skill, using approximate "cut offs"
for my 5 stages, based on average game lengths.

I thought I would also ask to see if anyone could
suggest better methods to establish more precise
cut offs but it looks like there aren't any easy ways.

I will start a new thread to propose my experiment.

MK

On Thursday, 15 December 2022 at 00:11:07 UTC, MK wrote:
> On December 14, 2022 at 6:48:44 AM UTC-7, Tim Chow wrote:
>
> > On 12/14/2022 7:40 AM, Axel Reichert wrote:
>
> >> MK <mu...@compuplus.net> writes:
>
> >>> Can we tell anything from the psition ID's?

I recently tried some quick and dirty experiments along these lines.

With the recent growing sophistication of AIs I had some hopes for automatic position ID classification (backgame, holding, racing cubes etc).

Unfortunately chatGPT and the new-ish Google Bard both seem to misinterpret (despite their confidence voices), even when directed to the URL with explicit instructions for gnubg ID encoding .

I have heard they perform much better at interpreting chess positions than backgammon positions. Maybe there was an element of internal human direction from the chatGPT Devs to improve performance because of the higher profile of chess.

IMO, it will need some more specific, focussed training to become proficient at interpreting and bulk classifying backgammon positions from IDs.

However, if this ever happens, I think it will be extremely useful.

Does anybody have any contacts in these companies?

On Saturday, April 1, 2023 at 2:56:41 AM UTC+1, sebalotek wrote:
> On Thursday, 15 December 2022 at 00:11:07 UTC, MK wrote:
> > On December 14, 2022 at 6:48:44 AM UTC-7, Tim Chow wrote:
> >
> > > On 12/14/2022 7:40 AM, Axel Reichert wrote:
> >
> > >> MK <mu...@compuplus.net> writes:
> >
> > >>> Can we tell anything from the psition ID's?
> I recently tried some quick and dirty experiments along these lines.
>
> With the recent growing sophistication of AIs I had some hopes for automatic position ID classification (backgame, holding, racing cubes etc).
>
> Unfortunately chatGPT and the new-ish Google Bard both seem to misinterpret (despite their confidence voices), even when directed to the URL with explicit instructions for gnubg ID encoding .
>
> I have heard they perform much better at interpreting chess positions than backgammon positions. Maybe there was an element of internal human direction from the chatGPT Devs to improve performance because of the higher profile of chess.
>
> IMO, it will need some more specific, focussed training to become proficient at interpreting and bulk classifying backgammon positions from IDs.
>
> However, if this ever happens, I think it will be extremely useful.
>
> Does anybody have any contacts in these companies?

In chess, the material balance already gives a lot of information about the type of position.
For example, if both sides have one rook and a king and the only other units present are pawns, then we know it's a rook ending.
In backgammon, the majority of games maintain a constant material balance until the no-contact phase.

A very crude undergrad-level software development assignment which mapped a chess position to its "type" using material balance only,
would not be wildly inaccurate. Of course, it wouldn't be able to reliably assess who (if any) had the better position.

Paul

On March 31, 2023 at 7:56:41 PM UTC-6, sebalotek wrote:

> On Thursday, 15 December 2022 at 00:11:07 UTC, MK wrote:

>> >>> Can we tell anything from the psition ID's?

> I recently tried some quick and dirty experiments
> along these lines. .....

How good to see that there was some continued,
silent interest on this topic/s from an old thread.

Also, as soon as I saw "sebalotek" I had a feeling
that I should remember something about you but
couldn't; until I searched RGB for the keywords
"sebalotek MK", which found only one older post
that contained (among many other things) a link
to the T-shirt you had created for me (which I still
have in my local backgammon folder... :)

http://i.imgur.com/axjPTgj.jpg

As I had clarified before, I had two ideas in mind:

1 - Plotting cube actions and cube values through
the various stages of games, out of my curiousity
to know when/where and how much the so-called
"cube skill" comes into play in gamblegammon.

2 - Improve on the very crude first "Murat mutant"
cube experiment (that Axel had done) using a still
very simple strategy (i.e. similar to mine, based on
"how much checker play is still left in the game"),
to not just do better than expected (compared to
the mutant's cube error rate) but to do better than
50% against the jackoffski formula.

I had drafted an article for a mutant strategy based
on cube actions depending on at which of the five
basic stages of the game but I never got around to
posting it. With my interest rekindled, I will try to do
it soon.

> IMO, it will need some more specific, focussed
> training to become proficient at interpreting and
> bulk classifying backgammon positions from IDs.
> However, if this ever happens, I think it will be
> extremely useful.

Can you elaborate on what you have in mind (i.e. in
what ways it will be useful, etc.)..??

MK

On December 14, 2022 at 6:48:44 AM UTC-7, Timothy Chow wrote:

> On 12/14/2022 7:40 AM, Axel Reichert wrote:

>> MK <mu...@compuplus.net> writes:

>>> Can we tell anything from the psition ID's?

>> Obviously, because it encodes the full state
>> of checkers on the board.

> Well, it's not possible to tell from the state of
> the checkers on the board how many moves
> have occurred since the start of the game.

Axel never replied to my request of elaborating
on what he had meant but I just thought about
it some more today.

Let's forget about deriving the stage of the game
from the position ID or the state of checkers on
the board and let's think about how else we can
make use out of those..?

What about the state of checkers on the board in
conjunction with the equity and/or GMC/MWC of
the position ID's..??

Any creative thoughts..???

MK

sebalotek schrieb am Samstag, 1. April 2023 um 03:56:41 UTC+2:

> With the recent growing sophistication of AIs I had some hopes for automatic position ID classification (backgame, holding, racing cubes etc).
Do we *really* need an AI to categorize a holding game, a backgame or a racing game?

And how do you want to train a bot on that? With supervized training you need a human to setup the training positions, which is tedious (we're not talking on a 3 digit number of examples) and with self learning (e.g. k-nearest or the like) it isn't clear what the categories will mean. I have tried such thing about 15 years ago, and what worked quite well on low dimensional problems, didn't work at all with BG, but naturally I could have coded rubbish and/or the HW at that time was to slow. But even then the categories might not make sense to humans (e.g. if you have a NN discriminating between cats and dogs, do you belive there are easy rules extractable from the net?)
> Unfortunately chatGPT and the new-ish Google Bard both seem to misinterpret (despite their confidence voices), even when directed to the URL with explicit instructions for gnubg ID encoding .
chatGpt more or less calculates the next word by probabilities and it is absolutely amazing how well that works, but you shouldn't make the error to believe that chatGPT knows anything

> I have heard they perform much better at interpreting chess positions than backgammon positions. Maybe there was an element of internal human direction from the chatGPT Devs to improve performance because of the higher profile of chess.
A modern chess program looks some dozen moves in the future, BG bots at most 10, having a good understanding even when looking just at the board. Which bots are then better in interpreting positions (I haven't looked deeply at Lc0 or the recent NN Stockfish, but it probably still holds true that 1 ply play is abysmal).

On April 2, 2023 at 4:27:09 PM UTC-6, Frank Berger wrote:

> sebalotek schrieb am 1. April 2023 um 03:56:41 UTC+2:

>> With the recent growing sophistication of AIs
>> I had some hopes for automatic position ID
>> classification (backgame, holding, racing cubes etc).

> Do we *really* need an AI to categorize a holding
> game, a backgame or a racing game?

I wondered that also.

> And how do you want to train a bot on that? .....
> But even then the categories might not make sense
> to humans (e.g. if you have a NN discriminating
> between cats and dogs, do you belive there are
> easy rules extractable from the net?)

What about not training the bots to categorize
positions but making them recognize certain
patterns, "spectrums" in series of positions from
analysing already played games for any possible
purposes as they had been discussed in these
long threads with many other participants:

https://groups.google.com/g/rec.games.backgammon/c/oxmW6YoyTsM/m/2dq6X1yoAQAJ

https://groups.google.com/g/rec.games.backgammon/c/Kxd9V4OA4PY/m/1mA7liVhCwAJ

https://groups.google.com/g/rec.games.backgammon/c/8bYqsL4BTjM/m/9QUkGqvnAgAJ

Perhaps you two hadn't participated in those
past discussions because the subject wasn't
interesting enough for you or the concept was
above the heads?? It's never too late... ;)

MK