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.
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).
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.
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?
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.
On 11/1/2022 7:48 PM, Timothy Chow wrote:
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 theIt won't be too long before AI will be able to do just that - adjust its
philosophy upon which these bots are based. If you want to
go down that road, then there are a million other changes to
consider.
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.
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.
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.
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.
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
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 queenI don't think this is historically accurate.
by taking longer than the required 50 moves.
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.
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.
Great chess players tend to play the board not the opponent
On 11/4/2022 12:59 PM, ah...Clem wrote:
Great chess players tend to play the board not the opponentI 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.
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 stronglyIt's not clear to me that playing a move which you know how to
resistant to moves that they see as objectively poor, even if such moves offer a lot
from a results-maximisation standpoint.
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
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.
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.
On 11/5/2022 10:04 AM, peps...@gmail.com wrote:
How good are you at chess?I never got an official rating, but I'd estimate my peak to
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.
have been around 1900 FIDE. But like you, I'm inactive now,
so I'm surely worse than that.
---
Tim Chow
Great! It seems a game between the two of us might well be a (somewhat) close contest.
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.
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.
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.
in bg my prediction that it will be based on generalized pattern
recognition
Any stats about the frequency of backgammon positions?
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?
Sysop: | Keyop |
---|---|
Location: | Huddersfield, West Yorkshire, UK |
Users: | 297 |
Nodes: | 16 (2 / 14) |
Uptime: | 14:25:23 |
Calls: | 6,667 |
Calls today: | 1 |
Files: | 12,216 |
Messages: | 5,336,564 |