I think XG should have a beavering threshold

[peps...@gmail.com]

Suppose XG is offered a double, and XG believes that,
after accepting, XG has a positive equity of 0.001.
Should XG beaver?
I think not -- at least some weight needs to be given
to the info that XG's opponent thought it was a double
(Kauder paradox positions are somewhat rare).

I think XG should use a positive beavering threshold where
it only beavers if it thinks its equity (after taking but before
beavering) is over that threshold. I would suggest a default
of 0.1, but that's quite arbitrary.

Paul

[ah....Clem]

I would expect XG to beaver whenever its equity increased by offering a
beaver. Anything else, such as a threshold, seems like sub-optimal
play. Why would I want that in a bot?

I say this having never been beavered by XG or anyone else.

--
Ah....Clem
The future is fun, the future is fair.

[Timothy Chow]

On 11/1/2022 4:12 PM, peps...@gmail.com wrote:

> Suppose XG is offered a double, and XG believes that,
> after accepting, XG has a positive equity of 0.001.
> Should XG beaver?
> I think not -- at least some weight needs to be given
> to the info that XG's opponent thought it was a double
> (Kauder paradox positions are somewhat rare).

Taking into account the opponent's behavior violates the
philosophy upon which these bots are based. If you want to
go down that road, then there are a million other changes to
consider.

---
Tim Chow

[ah....Clem]

It won't be too long before AI will be able to do just that - adjust its
play to exploit whatever your particular weakness might be.

The idea that AI is constrained to only play against another optimal
opponent is so last century.

Not sure if anyone's going to bother to train a neural net to do that
with backgammon, but they're already doing it with social media, which
is much more lucrative at scale than backgammon will ever be.

[MK]

On November 1, 2022 at 8:30:07 PM UTC-6, ah....Clem wrote:

> It won't be too long before AI will be able
> to do just that - adjust its play to exploit
> whatever your particular weakness might be.

Ha ha ha! There goes the ER/PR bulshit... :)

Ha ha ha! I'm having cramps in my groin from
laughing so hard... :))

> The idea that AI is constrained to only play
> against another optimal opponent is so last
> century.

I remember being the first and the only one
to ever argue that there can be more than just
one optimum strategy in backgammon 10-20
years ago. Was I ahead of my time..?

> Not sure if anyone's going to bother to train a
> neural net to do that with backgammon, but
> they're already doing it with social media, which
> is much more lucrative at scale than backgammon
> will ever be.

This is an insult to all those benevolent people
who contributed to Gnubg in the past and who
are still offering to help for a better bot for free
and to the charitable XG team who chose to sell
the best gamblegammon bot in the universe for
only $50 while they could sell it for $500...? ;)

Ha ha ha! I just farted from laughing so hard... :))

MK

[MK]

On November 1, 2022 at 5:40:32 PM UTC-6, ah....Clem wrote:

> I would expect XG to beaver whenever its equity
> increased by offering a beaver. Anything else,
> such as a threshold, seems like sub-optimal
> play. Why would I want that in a bot?

Because I happened to read and respond to your
next post, I can foresee that in less than 3 hours,
you will answer your own question by saying that
future bots will do similar (if not same) things to
take advantage of its opponents' weaknesses... ;)

Ha ha ha! This is farting funny... :))

MK

[peps...@gmail.com]

One problem that I find interesting in chess is beating an opponent
with a queen and a king against a rook and a king (no other pieces).

In a typical position, this is a very difficult problem indeed to solve
over the board without having learned the solution.
However, it's only difficult for a human. The number of possible positions
is less than 16 million, which makes it trivial to program a computer
to instantly provide optimal play for both sides.
However, such definitions of "optimality" are somewhat artificial.
Natural definitions involve taking the rook in the shortest number of moves
or (minimax) maximising the opponent's minimum solution.
This computer programming problem was solved in the 1970s if not earlier.
Without any systematic exposition of the solution, this endgame was so
difficult that famous grandmasters would sometimes draw with the queen
by taking longer than the required 50 moves.

There was talk, though, about rewriting the code to take human psychology into
account -- for example, the computer could create positions where all winning paths
involved unnatural-looking moves.
However, I've never seen any research on Q v R from a bot v human perspective.
It would be interesting to see whether an anti-human program could draw against
a GM in this ending.
Currently, I think the great majority of GMs have studied Nunn's write-up of this ending,
and can therefore comfortably beat the standard bot tablebase algorithms for this ending.

Paul

[MK]

On November 1, 2022 at 5:48:16 PM UTC-6, Tim Chow wrote:

> On 11/1/2022 4:12 PM, peps...@gmail.com wrote:

>> ... at least some weight needs to be given to the

>> info that XG's opponent thought it was a double

> Taking into account the opponent's behavior
> violates the philosophy upon which these bots
> are based. If you want to go down that road, then
> there are a million other changes to consider.

While resting and recovering from a stiched up
groanial hernia that I suffered from laughing too
hard last night, I have some more to say about
future AI bg bots, hoping that you and others will
contribute to also.

Starting with your above response to Paul's idea,
current bots have already been "improved" with
so many similar, arbitrary tweaks and patches
violating the initial philsophy, which was faulty to
begin with, that they became mixed garbage piles
with very little left that can be recovered/recycled.

Just as an example from a recent post by Phillip:
"... backgammons (in non-contact positions) are
computed explicitly instead of using the neural net.
Moreover there is a serie of sanity checks to detect
certain win / certain loss / certain gammon /
impossible gammon situations and override the
neural net outputs if needed".

These started with TD-Gammon v.2 and infected all
current bots descended from it. In fact, I've always
opposed to calling them AI bots. If you don't, surely
Frank would remember well, since he had defended
that his bot was AI depending on what's meant by it.

I've also always predictingly argued that if and when
a truely AI bg bots comes along, it will truen upside
down you people's fantasy world of gamblegammon,
by debunking all that bullshit "cube skill theory", etc.

I had proposed that the first generation of truly AI gg
(gamblegammon) bots would be self-trained through
cubeful random play from the beginnings, instead of
formulating "cube skill" into cubeless play afterwards.

I had then proposed that the second generation of
truly AI gg bots would perpetually keep self-learning
through non-random play (i.e. actual games against
opponents other than itself) to the point of developing
different strategies and deploying them alternatively
as best to exploit their opponents.

It seems like some of you are beginning to understand
my arguments even if skipping the first generation and
jumping to the second generation, which is fine since
they can be achieved together.

The first generation AI bots would debunk the current
fallacious skill theories and faulty equity calculations
but would still be only capable of single (i.e. optimum)
strategy.

The second generation AI bots capable of multiple,
adaptable strategies would completely demolish all
of the remaining bullshit like ER/PR since they can
only exist if there is just a single, optimum strategy.

Without the limitation of a single, optimum strategy
many other dominoes will also. For example, claims
that Hypergammon is solved and can be played, by
bots like Gnubg, perfectly will have to be retracted.
A Hypergammon bot that can adjust its game to its
opponent will become unsolvable.

The same goes for Tim who claimed to have solved
Murat's Hypestgammon mathematically. He never
compared the accuracy of his calculations against
a cubefully-trained Hypestgammon bot but that may
become moot, as the challenge will get bigger. Even
such a simple gg variant with zero checker skill may
become much harder, if not impossible, to solve if the
Hypestgammon bot can adjust its minimalistic cube
strategy according to its opponent.

Let me finish this long post by saying that human
bg and gg players have always been "unsolvable"
and able to adjust their strategies to their opponents.

I'm sure quite a number of humans players have the
ability to consistently beat the best of the current
bots even if they may not be aware of their abilities.

I'm impatiently waiting for bots to catch up to human
players soon. And then we can progress on to talking
about bots that will be unbeatable by humans...

MK

[Timothy Chow]

On 11/1/2022 10:30 PM, ah....Clem wrote:
> It won't be too long before AI will be able to do just that - adjust its
> play to exploit whatever your particular weakness might be.

In principle, the technology should exist today. But it seems that
there's not enough incentive for anyone to bother with programming a
backgammon bot in this way.

---
Tim Chow

[Timothy Chow]

On 11/2/2022 9:06 AM, peps...@gmail.com wrote:
> One problem that I find interesting in chess is beating an opponent
> with a queen and a king against a rook and a king (no other pieces).
>

[...]

> This computer programming problem was solved in the 1970s if not earlier.
> Without any systematic exposition of the solution, this endgame was so
> difficult that famous grandmasters would sometimes draw with the queen
> by taking longer than the required 50 moves.

I don't think this is historically accurate.

The watershed moment was Walter Browne versus Belle, 1978. Prior to
that, this was considered an "easy" win for the side with the queen.
I'm not aware of any human vs. human GM games where the 50-move limit
was overstepped.

Belle breathed new life into the ending by demonstrating a new defensive
technique for the side with the rook. When Browne first encountered
this new defense, he indeed overstepped the 50-move limit. But then he
thought about it for a while (without computer help), and figured it
out, beating Belle on his next attempt.

There could be some GM's today who haven't mastered this ending, but I
would be surprised.

Anyway, even if K+Q vs. K+R isn't a good example to illustrate your
point, we could pick some other ending that human GM's haven't all
mastered. Still, I'm not sure that such an ending would provide a
good test case for your experiment, because I suspect that the trickiest
moves for a human to counter are almost always the moves that take the
longest to win against. That is, there's not much difference between
the traditional "best defense" and the "most stubborn defense against
a human" in these tablebase endgames.

I have heard that some engines have something called a "contempt factor"
that causes them to try to take advantage of an inferior opponent, but
I'm not familiar with the details.

---
Tim Chow

[MK]

On November 2, 2022 at 10:09:10 PM UTC-6, Tim Chow wrote:

> On 11/1/2022 10:30 PM, ah....Clem wrote:

>> It won't be too long before AI will be able to do
>> just that - adjust its play to exploit whatever
>> your particular weakness might be.

> In principle, the technology should exist today.

Processing power and storage capacity already
exist, as does programming knowledge needed.

> But it seems that there's not enough incentive

The reason for this may deserve expolring. The
philosophy on which the current bots are based,
that you referred to in another post, has become
too dogmatic now and the flock is content with it.

Thus, anyone who would want to develop such a
bot, would have to go against the entire existing
gamblegammon community and establishment.

This would be true even if the existing bots could
be remodeled without any philosophical changes
because "adjusting one's play", (i.e. "no optimum
play"), will make it impossible to tell if a checker
or cube decision is made in error or on purpose in
order to exploit the opponent's weakness.

And once you can't discern errors, then the error
rate concept will go out the window entirely. Then
player ratings will become based on actual game
winning performance, (more like in chess, which
will actually be a good thing, as using ELO ratings
in bg will become more meaningful).

When there are efforts to introduce error ratings
into chess, (which many people, including myself,
find totally nonsensical), measured in bullshit like
millipawns, etc. I doubt that the gamblegammon
world will progress in the opposite direction by
rehabbing from their obsessive/compulsive low
ER/PR disorder and embrace new ideas or bots...

MK

[peps...@gmail.com]

Hi Tim,

Thanks for the details.
I meant the Browne v Belle events. When I talked about human difficulty,
I meant that it was difficult for humans to beat the bots until Nunn explained the method but I don't think
I said what I meant.
What I'm saying now may also be wrong -- this thing about Nunn explaining it being the
main factor in human understanding might be my imagination combined with
over-reverence.

I have read (but not in full detail) Nunn's work on K + Q vs K + R.
This particular ending has the characteristic that the computer has
lots of opportunities to give the human a slightly shorter win in order
to force the human to play counter-intuitively.
That's why, in Nunn's text, he often recommends lines that are 1 or 2 moves
longer than optimal.

Re being surprised about GMs, the thing is that there are so many of them.
Maybe, in direct contradiction to what you say, the surprise would be if every
single one of them has mastered the ending.

It may be a pedantic quibble, but I personally would be surprised if few than 95%
of them haven't mastered the ending, and I would also be surprised if more than
1 of the top ranked 100 GM's haven't done so.

As an afterthought, you may mean that, if you met a GM who didn't know this
ending well, you'd be surprised. I'd be surprised at this, too.

Paul

[ah...Clem]

I think you're right about that. It's hard enough to get Xavier to work
on XG.

I think poker is a more likely candidate, since so much depends on what
your opponents do, there's a lot of hidden information, and it's a lot
more popular.

Great chess players tend to play the board not the opponent, great poker
players learn how their opponents play and act accordingly. Not sure if
the current poker bots keep a database of opponents habits, but it would
seem to be a measurable advantage. If YouTube can tailor it's offerings
to each user, a poker bot would be able to do the same.

[Timothy Chow]

On 11/4/2022 12:59 PM, ah...Clem wrote:

> Great chess players tend to play the board not the opponent

I once read an interesting article, by Andy Soltis I think, who
was of the opinion that there were great chess players of both
types (play the board vs. play the opponent). I believe that
Geller was known for trying to play the optimal move no matter
who he was playing, whereas Emmanuel Lasker was famous for
adjusting his play to his opponent.

Nowadays, I feel pretty confident that all the top players
"play the opponent" at least in the opening phase, meaning that
they carefully study all the available data about their upcoming
opponent's opening choices, and plan their own opening prep
accordingly. After the opening phase, I expect that different
players will "adjust to the opponent" to different degrees.

There's another sense in which one can "play the opponent":
especially in positions where you're worse, it often makes sense
to play moves that create complications even if in some sense
they are not "objectively best." Here, one is not necessarily
making Play A against Opponent X and Play B against Opponent Y;
one might make the same play against all (sufficiently strong)
opponents, but the play is dictated more by its practical chances
than by the objective demands of the position. Back in the day,
Mikhail Tal would infuriate some of his rivals by playing "unsound"
moves and winning anyway. I'm sure there are top players today
whose play could be described in this way, especially in rapid or
blitz chess.

---
Tim Chow

[peps...@gmail.com]

I can't prove this but I sense that ah..Clem has a point in the following sense:
To simplify, let us consider two types of thinking: "board thinking" where you think
about the objectively best chess move from a very pure standpoint and
"opponent-oriented thinking" where you think about the particular opponent.
My (perhaps controversial) point, which may be what ah..Clem's point is, is that, from
a rational and performance-maximisation standpoint, there is generally speaking,
among the best players, far too little "opponent-oriented thinking". For example,
even Tal probably never (or almost never) played a deliberately unsound sacrifice
(except perhaps out of desperation in a worse position). The point about Tal
was never deliberate unsoundness, but rather that he was prepared to make sacrifices
even if he was unable to check whether they were sound or not. Those sacrifices of
Tal that were unsound, were discovered to be so during post-mortems, not OTB.
I think that, if Tal saw a refutation to a sacrifice, he wouldn't play it even if he thought
his opponent would be unlikely to spot the refutation.

Great chess players tend to love chess (of course) and therefore tend to be strongly
resistant to moves that they see as objectively poor, even if such moves offer a lot
from a results-maximisation standpoint.

Paul

[Timothy Chow]

On 11/5/2022 7:46 AM, peps...@gmail.com wrote:
> I think that, if Tal saw a refutation to a sacrifice, he wouldn't play it even if he thought
> his opponent would be unlikely to spot the refutation.
>
> Great chess players tend to love chess (of course) and therefore tend to be strongly
> resistant to moves that they see as objectively poor, even if such moves offer a lot
> from a results-maximisation standpoint.

It's not clear to me that playing a move which you know how to
refute (except out of desperation, as you said) is going to be
a results-maximizing move very often, for a top player. If the
opponent is much weaker, then the top player doesn't need to
take risks of this sort, but will almost surely win anyway. If
the opponent is equally strong or almost equally strong, then
the risk is high that the opponent will see the refutation that
you saw.

The one case where I might agree with you is in world championship
matches. It does seem to me that the players are so terrified of
losing a game that they shy away from even reasonable risks. I'm
thinking for example of the final classical game in the Caruana-
Carlsen match. Carlsen, in my opinion, did the chess equivalent
of cashing while too good, by offering a draw in a position where
he could have at least played for a win for a few more moves, with
no risk. Several commentators made this point (and not just bozos
like me, but top players like Anand).

---
Tim Chow

[peps...@gmail.com]

How good are you at chess?
At my peak, I was 1700 USCF (actually, it may have been something like 1696)
but I'm completely inactive now so I'm probably somewhere in the 1500 to 1600 range.

Paul

[Timothy Chow]

On 11/5/2022 10:04 AM, peps...@gmail.com wrote:
> How good are you at chess?
> At my peak, I was 1700 USCF (actually, it may have been something like 1696)
> but I'm completely inactive now so I'm probably somewhere in the 1500 to 1600 range.

I never got an official rating, but I'd estimate my peak to
have been around 1900 FIDE. But like you, I'm inactive now,
so I'm surely worse than that.

---
Tim Chow

[peps...@gmail.com]

Great! It seems a game between the two of us might well be a (somewhat) close contest.

Paul

[Timothy Chow]

On 11/5/2022 10:13 AM, peps...@gmail.com wrote:
> Great! It seems a game between the two of us might well be a (somewhat) close contest.

Oh, by the way, here's a somewhat more recent data point.
A few years ago, I got hooked on Puzzle Rush for a few months.
I recall that my goal was to score 30, but I never quite managed
to do it---I think my top score was 28. Maybe if I had paid for
a subscription so that I could play it more than a couple of times
a day, then I would have hit 30 eventually.

Of course, Puzzle Rush measures only a very narrow sliver of chess
skill, but it's more than nothing.

---
Tim Chow

[peps...@gmail.com]

I'm much better at chess problem solving than playing.
For example, I got a congratulation in this article:
https://www.schakeninhoogland.nl/2021/12/31/sietske-greeuw-wint-oloth-2021/ and I actually
bettered this performance by coming 3rd more recently.

Paul

[MK]

On November 4, 2022 at 10:59:24 AM UTC-6, ah...Clem wrote:

> On 11/3/2022 12:09 AM, Timothy Chow wrote:

>> On 11/1/2022 10:30 PM, ah....Clem wrote:

>>> It won't be too long before AI will be able to do
>>> just that - adjust its play to exploit whatever
>>> your particular weakness might be.

Just ignore me, guys. In fact, don't even read past
this. I'm just posting to create a record for future.

The words "sacrifice" and its derivatives have been
used in RGB since way back when initially referring
to giving up points, blots, etc.

Then people started to talk about sacrificing equity
interchangeably with giving up, losing equity.

I believe the first time it was used (6?? years ago)
in relation to PR was about some "inferior" moves
Mochi had made against other humans (I believe it
was Paul while arguing that such a giant must have
made moves sacrificing PR knowingly, in order to
exploit his (assumedly inferior) opponents.

In time, exploiting one's opponent by such moves
has become common talk and acceptable as long
as the player pointed out his purposefully inferior
moves before making them.

At one time, Stick had dared me to play against him
for money instead of me playing against XG. I saw
no point in playing against an unpredictable human
with a higher PR than XG but I said I would do it with
the condition that if he strayed to far from his PR by
makin PR-sacrificing moves, (i.e. adjusting his game
to me), he would be proportionately punished.

About 4 years ago, I was the first one to start talking
about exploiting bots, (assumedly superior to even
human giants), using the words as "PR-sacrificing"
format and also coining the expression "PR-defying".

With it, I also started to argue more confidently and
more avidly that multiple "optimum" strategies were
possible in bg, in lengthy discussions with Tim, et al.

One weakness of the bots was their being perfectly
consistent thus completely predictable, that humans
could exploit by "steering" the bots by "bot-defying"
moves but that bots weren't capable of doing the same.

However, I had predicted that future, truly AI (similar
to Alpha-Zero) bots would drop the "single optimum"
strategy and would be able to adjust their strategies
to their opponents.

Now, you should be able to understand why I just won't
let some bozo/s come plagiarize my pioneering ideas
after having derided them for years in the past.

As the saying goes, "he who laughs last, laughs best".
And yes, it's now my turn to laugh my ass off to you all
even at the risk of suffering groanial hernies... :)

>> In principle, the technology should exist today.

> Great chess players tend to play the board....
> great poker players learn how their opponents play....

> Not sure if the current poker bots keep a database
> of opponents habits, but it would seem to be a
> measurable advantage.

I don't know much about chess or poker but in bg my
prediction that it will be based on generalized pattern
recognition, similar to what I had proposed in the past:

"Trying to understand what you wrote, I thought
"of color spectrums of elements. I wonder if the
"graphs of bot-v-bot, human-v-human, human-v-bot
"positions graphs can be similarly recognizable
"enough characteristic? What about even subsets
"like strong-bot-v-weak-human, weak-human-v-weak
"-human, weak-bot-v-strong-human, etc...?? Can we
"look at a "spectrum of positions" and be able to say
"that those come from games played by weak-bots
"and strong-humans?

Here is the link if you want to read the entire thread:

Any stats about the frequency of backgammon positions?

https://groups.google.com/g/rec.games.backgammon/c/oxmW6YoyTsM/m/Hxu16HZFCgAJ

An enormous database of all players is unnecessary.
The bots will need to be able to play off-line, without
needing know their unique opponents (by requiring
them to identify themselves, which they may not do
or falsify), but by assesing their opponents during the
course of the actual games, matches.

It will work similar to face recognition, for example.
But since we won't be trying to identify criminals, etc.
recognizing general facial characteristics will be good
enough for the purpsoe. At the top layer may be race,
i.e. asian, african, eurepean, etc. Then sex, i.e. female,
male. Then age, i.e. young, middle-aged, old. Etc...

Future bots' ability to deploy alternative, conditional
strategies will only be the icing on the cake. The first
priority needs to be achieving the most important step
of creating bots without human bias, such as bullshit
theories of skill, founded on arbitrary formulas, etc.

I wish future was now...

MK

[Axel Reichert]

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

> in bg my prediction that it will be based on generalized pattern
> recognition

Patterns recognized such as "pointless doubles"?

> Any stats about the frequency of backgammon positions?

For example, frequencies of doubles given based on the number of home
board points?

Axel

[MK]

On November 6, 2022 at 2:19:14 AM UTC-7, Axel Reichert wrote:

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

>> in bg my prediction that it will be based on
>> generalized pattern recognition

> Patterns recognized such as "pointless doubles"?

Bots won't judge any decisions, including doubles,
subjectively as "pointful, "pointless", etc.

>> Any stats about the frequency of backgammon
>> positions?

> For example, frequencies of doubles given
> based on the number of home board points?

"Doubles" aren't "positions". I wan't talking about
decisions.

MK