Adding this complexity and mental overhead .....
..... this reminds me on Ptolemy's epicycles. (-;
My gut feeling says that any major improvement
on race cube decisions will not be based on adding
this feature or that ..... but rather on something
fundamentally different, namely EPCs along with a
doubling criterion matched to them.
Even if algo X is the best when used in an automated
botlike fashion ..... this by no means indicates that
algo X is the best for the informal approach. I think
Stick's method is [Keith Count + informal intuitive
adjustments] There's no evidence that Isight beats
this (or matches this) .....
through my experiments playing against the bots, I have shown that
other/new approaches could work better
I can't wait for the day you will liken cube skill, equity, luck,
MET's, ER/PR, etc. calculations to Ptolemy's epicycles
Did I see the words "informal intuitive adjustments"? And to say, no
less than, their being better than "algos used in an automated botlike fashion"?? Wow!
my beating the bots didn't mean that I was the greatest backgammon
player on the planet but it just meant that the bots weren't as strong
as I am
first is an argument against beavers. ...
a beaver functions (among other things) as a sharp rebuke of the
opponent's play -- "Your double was so inaccurate!
particularly if we also have automatic doubles.
The beavered player might have positive equity but you can't
necessarily show that in practice.
"peps...@gmail.com" <peps...@gmail.com> writes:
a beaver functions (among other things) as a sharp rebuke of theYes. A double allows only the player on roll (who is "on the brink of
opponent's play -- "Your double was so inaccurate!
the unknown") to raise the stakes, while a beaver allows both players to "comment" on the same position (in case of further critters, multiple
times).
This is a significant difference (also from a procedural point of view,
the beavering player is not on roll) to a simple take and redoubling on
next turn: Except when neither player could move, the position will be a different one.
The exception gives me the idea that it could be interesting to think
about an "eternal beaver" position similar to
https://www.bkgm.com/rgb/rgb.cgi?view+366
that might be both a correct double and a correct beaver. Then the
stakes would explode even more rapidly ...
particularly if we also have automatic doubles.Just a highly volatile equivalent of multiplying the stakes with
1.25. No need for this.
The beavered player might have positive equity but you can't... because of the possibility of a Petersburg paradox. Enough reasons
necessarily show that in practice.
to get rid of it.
1. Normal checker play, but mutant cube strategy to jack the cube up[...]
and fabricate a Petersburg paradox, which drowns the signal in the
noise.
P. S.: Both strategies are no argument against the doubling cube. The
first is an argument against beavers.
The article you link to is indeed interesting but very well-known. It
shows (or claims to show) the concept of a position without a
well-defined equity.
In the position you link to, beavering would be a massive mistake --
clearly the player on roll has a big advantage.
Beavers were pretty clearly invented by gamblers.
Timothy Chow <tchow12000@yahoo.com> writes:
Beavers were pretty clearly invented by gamblers.
Is there more known about the history?
On 4/24/2022 9:29 AM, Axel Reichert wrote:
Timothy Chow <tchow12000@yahoo.com> writes:
Beavers were pretty clearly invented by gamblers.Is there more known about the history?
Some standard links for the history of the doubling cube are:
http://www.chicagopoint.com/bgdoubling.html
https://bkgm.com/articles/Morawski/30sUnderstandingOfDoubling.html
tacking "[citation needed]" on to "Prostitution is one
of the oldest professions."
MK <mu...@compuplus.net> writes:
through my experiments playing against the bots, I have
shown that other/new approaches could work better
If you are referring to our "mutant" discussion: You just
found a way that ends up as a Petersburg paradox and
thus makes quantification more difficult in a real live
session, but still possible in an analytical way. This cube
strategy is vastly inferior.
0% was just "a" Murat mutant, the minimum I didsettle for just not lose the chance of making you do the
Did I see the words "informal intuitive adjustments"?
And to say, no less than, their being better than "algos
used in an automated botlike fashion"?? Wow!
Do not mix things up. Racing double algorithms such as
my Isight method are crutches for mere humans.
my beating the bots didn't mean that I was the greatest
backgammon player on the planet but it just meant that
the bots weren't as strong as I am
My suggestion is that you do 100 1-point matches against
the bot of your choice and report the results.
1. Normal checker play, but mutant cube strategy to jack
the cube up and fabricate a Petersburg paradox, which
drowns the signal in the noise.
2. Normal cube strategy, but mutant checker play aiming
at ultra-deep backgames or primes in the opponent's
outfield (which the neural nets have seldomly seen and
thus might get wrong) will result in a similar effect, because
then the bot will use a "mutant" cube strategy.
My prediction is you will get trounced. Good luck!
P. S.: Both strategies are no argument against the doubling
cube. The first is an argument against beavers.
The second in my opinion is not even an argument for your
imagined "AlphaGammon", because I expect that, left on
their own, bots "from scratch" will not see these positions
frequently enough.
I've made the point before that I think that beavers- without-allowing-raccoons is wrong.
... because of the possibility of a Petersburg paradox.
Enough reasons to get rid of it.
On April 24, 2022 at 1:53:16 AM UTC-6, Axel Reichert wrote:
[Mutant] cube strategy is vastly inferior.
You have no justification to make this last ass-ertion.
Your experiment as I had proposed it, ended after the first set of
real life games. After that it was "smoke and math" or "maths and
mirrors"...
Have you looked at my backgammon website
The "Murat mutant" that you used in your experiment is not Murat!
Let's call it "murat-mutant-a". It was a way for me to make you to
prove it to yourself that the "cube skill theory" is fancyfully
elaborate bullshit. You demonstrated that.
I don't give a rat's ass about Petersburg paradox, etc.
Take a look at my published experiments before you ask more from me.
critters that follow raccoons. They should all be allowed. Or, in the alternative, a mathematical proof needs to offered that the so-called
cube skill stops at raccoons or even before that at beavers.
In the "Murat mutant" experiment, I'm not sure how Axel actually
handled this.
MK <mu...@compuplus.net> writes:
You have no justification to make this last ass-ertion.
I do, I did the maths. Feel free to point to any errors in
my derivation.
Your experiment as I had proposed it, ended after the
first set of real life games. After that it was "smoke and
math" or "maths and mirrors"...
Which numbers are wrong? In which lines are errors?
The "Murat mutant" that you used in your experiment is
not Murat! Let's call it "murat-mutant-a".
It was a way for me to make you to prove it to yourself
that the "cube skill theory" is fancyfully elaborate bullshit.
You demonstrated that.
To remind you (quoting here):
By the way, in 10000 games with 1 beaver allowed, double > 0.5
and take > 0.0 the mutant lost 62117 against gnubg's 84870.
I would rather say that both the session and the math that
built upon it (showing the mutant strategy is expected to
lose 0.7 cube-normalized points per game in the long run
against GNU Backgammon) are pretty strong arguments
AGAINST your claim.
Take a look at my published experiments
I did so. The money sessions were done with beavers and
raccoons, so suffer from the Petersburg paradox (you bring
home a lucky game with lots of critters and protect your
lead from then on).
The matches obviously cannot suffer from the Petersburg
paradox, so a different explanation is needed: In almost all
matches (and I checked all your published 5-point matches,
15-point matches, and 25-point matches) that you won you
were considerably luckier than GNU Backgammon.
Do a 100 games money session without beavers, and do
100 1-point matches.
You first ran 1,000 games. Then 5,000 without beavers,
3,000 with beavers and 3,000 with raccoons. Later on in
discussion we all kept talking about 10,000 games but I
never understood if that was a subset of the total of the
above numbers of if was a new/different set of games.
When you started the experiment, you admitted that you didn't really understan Markov chains. Then you went on to derive 5 billion (or
more?) games from that 10,000.
Is there even a point in looking for errors in such a Mickey Mouse experiment?
If we conduct more experiments later with combinations like doubling
at 55% and taking at 5%, doubling at 45% and taking at 15%, etc. we
can call those "mutant-b", "mutant-c", "mutant-d", etc.
a mutant's goal is not to replace it with another "theory" but just to
poke a hole in it as a first step
Can you please clarify what was my claim or how you had understood it?
This is a crudest possible monkey wrench experiment
That's the results of the fallacious luck calculations by the bot.
Do a 100 games money session without beavers, and do
100 1-point matches.
What would be the point of this? To reduce lucky high cubes?
I am the one who argues that cube magnifies luck, remember?
MK <mu...@compuplus.net> writes:
O. K., fair point, since I kept adding incrementally
to my session. Overall, I had 10000 games (with
beavers, but no raccoons allowed).
I did some independent short sessions with
raccoons, but later found out two things:
1. They are not needed, since it is trivial to factor
them in. "Mutant-a" doubles above 50 %, so it will
kind of automatically raccoon if beavered.
2. GNU Backgammon has a cube limit of 4096, so
allowing raccoons made it far more likely to have a
game violating this limit and thus skew the results
of the session.
So no need to allow raccoons, just assume automatic
ones.
When you started the experiment, you admitted that you
didn't really understan Markov chains. Then you went on
to derive 5 billion (or more?) games from that 10,000.
Some people are able to learn. (-:
The Markov chain stuff was just helpful to confirm my gut
feeling about beavers and beyond leading to Petersburg
paradox. It is not needed for my (analytical) argument, but
it served to turn my head into the right direction.
That is how research often works. (-:
Is there even a point in looking for errors in such a
Mickey Mouse experiment?
"If in doubt or without clue, insult"? Where are your
Donald Duck experiments?
If we conduct more experiments later with combinations
like doubling at 55% and taking at 5%, doubling at 45%
and taking at 15%, etc. we can call those "mutant-b",
"mutant-c", "mutant-d", etc.
"We"? Not me, I am happy with the results of my research.
Go ahead if you feel more is needed.
Can you please clarify what was my claim or how you
had understood it?
If you only would know your claim yourself ...
Candidate 1: Cube theory does not matter
Easily disproved by rolling a dice for doubling decisions:
1, 2, 3 = Double, 4, 5, 6 = Hold. Likewise for taking decisions:
1, 2 = Beaver, 3, 4 = Take, 5, 6 = Pass. If cube skill does not
exist, this should give the same results as GNU Backgammon
playing itself. It does not, I did this, but was of course insulted
by you for carrying out this "meaningless experiment".
Candidate 2: Mutant cube handling is better than the bot's
Disproved by my 10000 games session and my subsequent
analytical work. The latter was necessary due to beavers
and beyond yielding a Petersburg paradox, which does not
occur without beavers.
Try a long session with only doubles, but no beavers to
convince yourself.
Candidate 3: Mutant checker play is better than the bot's
Disproved by doing 100 1-point matches. Your job (I
cannot algorithmically mimic your brain ...), not mine.
Sigh. There is is cube skill, see candidate 1, but there is
the Petersburg paradox with beavers and beyond. So the
"skill maximum" is with doubles only.
In almost all matches (and I checked all your published
5-point matches, 15-point matches, and 25-point matches) that
you won you were considerably luckier than GNU Backgammon.
In almost all matches (and I checked all your
published 5-point matches, 15-point matches, and 25-point matches) that
you won you were considerably luckier than GNU Backgammon.
On 4/30/2022 9:26 AM, Axel Reichert wrote:
In almost all matches (and I checked all yourThis fact by itself doesn't mean much. Take any single match between
published 5-point matches, 15-point matches, and 25-point matches) that
you won you were considerably luckier than GNU Backgammon.
any players, and with high probability the winner will also have
the higher luck rating.
I recall that someone took a bunch of Murat's money games and calculated
the total luck over all games of both players, and the luck was about
the same for both players.
One could do the same calculation for his matches, but to be meaningful,
the luck would have to be measured in terms of MWC rather than EMG.
On May 1, 2022 at 2:07:51 AM UTC-6, Axel Reichert wrote:
will I be right to understand that Gnubg will never double with its
MVC < 50% and will also never beaver with its MVC < 50%?
2. GNU Backgammon has a cube limit of 4096, so
allowing raccoons made it far more likely to have a
game violating this limit and thus skew the results
of the session.
No. It's you who skewed the results by not allowing!
Cube going sky high doesn't mean Petersburg Paradox.
Do you think you can tear of a page from a book and call
the rest still good?
Candidate 3: Mutant checker play is better than the bot's
Disproved by doing 100 1-point matches. Your job (I
cannot algorithmically mimic your brain ...), not mine.
What mutant? What checker play? What are you talking
about?
MK <mu...@compuplus.net> writes:
will I be right to understand that Gnubg will never double with
its MVC < 50% and will also never beaver with its MVC < 50%?
As a crude first approximation, yes.
2. GNU Backgammon has a cube limit of 4096, so allowing
raccoons made it far more likely to have a game violating
this limit and thus skew the results of the session.
No. It's you who skewed the results by not allowing!
No, but to understand that, you would need to accept or
learn some math, and I am pessimistic of you doing either. (-:
Cube going sky high doesn't mean Petersburg Paradox.
True, it is a necessary, but not sufficient condition.
But I checked the sufficient one (divergent geometrical series)
and showed it holds true when beavers are allowed and a
mutant cube strategy is used.
Do you think you can tear of a page from a book and call
the rest still good?
Yes, roughly.
Candidate 3: Mutant checker play is better than the bot's
Disproved by doing 100 1-point matches. Your job (I
cannot algorithmically mimic your brain ...), not mine.
What mutant? What checker play? What are you talking
about?
I ask YOU to do this: Play 100 matches to 1 point and report
the results. This will eliminate candidate 3.
MK <mu...@compuplus.net> writes:
[lots of vague/irrelevant stuff]
There is no value for me in continuing a fruitless discussion.
In case an interesting idea from you comes up, I will happily
(and perhaps even gratefully) pick it up, but otherwise spend
my time in genuine research rather than sophistry.
Sysop: | Keyop |
---|---|
Location: | Huddersfield, West Yorkshire, UK |
Users: | 296 |
Nodes: | 16 (2 / 14) |
Uptime: | 77:32:24 |
Calls: | 6,658 |
Calls today: | 4 |
Files: | 12,203 |
Messages: | 5,332,834 |
Posted today: | 1 |