It occurred to me (after like 5 seconds thought) that it would be fun to see
Axel employ his algo against the world's best humans (assuming some of
them are accessible).
My concept is this. Start with the longest possible symmetrical race --
this means each player has 15 checkers on on their 12 point -- and
go ahead using the normal rules of backgammon, treating the above
position as the initial position.
Would such a test be fair, or would there be a legitimate objection that
Isight is trained on real positions and, with my artificial opening position,
there's no guarantee that the positions would be representative of
normal play?
I think the objection stands, but it might make an interesting match.

Paul

"peps...@gmail.com" <pepstein5@gmail.com> writes:

> Start with the longest possible symmetrical race -- this means each
> player has 15 checkers on on their 12 point -- and go ahead using the
> normal rules of backgammon, treating the above position as the initial
> position.

[...]

> Would such a test be fair, or would there be a legitimate objection
> that Isight is trained on real positions and, with my artificial
> opening position, there's no guarantee that the positions would be
> representative of normal play?

Not only my Isight method is trained on real positions, but humans and
bots are as well. GNU Backgammon evaluates the position as follows:

0-ply: No double, take (12.6 % winning chances)
1-ply: Double, pass (88.8 % winning chances)
2-ply: No double, take (30.7 % winning chances)
3-ply: Double, pass (84.0 % winning chances)
4-ply: No double take (35.8 % winning chances)

Of course, even a weak (but untruncated) roll-out gives the correct
result that the player on roll is a slight favorite.

To my surprise, the checker play of the bot was weird in the beginning
as well (before the game reached more familiar positions), resulting in
a lot of unnecessary wastage. Together with the bot's premature doubles
on odd-ply settings (easy takes, and now the cube is on my side) it
might well be that a human has a good chance to make this a winning
proposition against the bot.

I will play around a bit, thanks for the idea!

Best regards

Axel

On September 5, 2023 at 2:08:10 AM UTC-6, Axel Reichert wrote:

> "peps...@gmail.com" <peps...@gmail.com> writes:

>> Start with the longest possible symmetrical
>> race -- this means each player has 15 checkers
>> on on their 12 point

>> Would such a test be fair, or would there be a
>> legitimate objection that Isight is trained on
>> real positions

I consider Paul a member of the dog pack who
won't bite his ilk but every so often he will pose
a question that may be considered as nibbling
at their heels like a cute puppy... :) and he never
pursues issues he raises to the point of causing
any real damage to his ilks... ;)

What I find also interesting here is the concept
of "real positions". A human could never create
a position like his example on purpose but could
create other positions unfamiliar to bots, (which
become idols of good sex through masturbation
and use of dildos like cube skill formulas applied
to cubeless equities), in order to beat them bots.

Maybe humans like me should consider those as
"positions unfair to the bots" also..?

> Not only my Isight method is trained on real
> positions, but humans and bots are as well.
> GNU Backgammon evaluates the position as
> follows:

> 0-ply: No double, take (12.6 % winning chances)
> 1-ply: Double, pass (88.8 % winning chances)
> 2-ply: No double, take (30.7 % winning chances)
> 3-ply: Double, pass (84.0 % winning chances)
> 4-ply: No double take (35.8 % winning chances)
>
> Of course, even a weak (but untruncated) roll-out
> gives the correct result that the player on roll is a
> slight favorite.

Axel, why do you waste your life fabricating excuses
to deny realities, including of your own findings..?

> To my surprise, the checker play of the bot was
> weird in the beginning....

All that bla bla bla just to say that Noo-BG is a "cow
cake"!

> I will play around a bit, thanks for the idea!

Have you folks noticed that I have been improving
my tactful language skills for a while now..? (i.e.
rooster eaters instead of cocksuckers, chocolate
smoothie slurpers instead of shit lickers, etc...

And so, "cow cake" instead of "bullshit".. :) And I
guess Axel's Isight would be the "Icing" on it...! ;)

MK

On Tuesday, September 5, 2023 at 4:08:10 AM UTC-4, Axel Reichert wrote:
> "peps...@gmail.com" <peps...@gmail.com> writes:
>
> > Start with the longest possible symmetrical race -- this means each
> > player has 15 checkers on on their 12 point -- and go ahead using the
> > normal rules of backgammon, treating the above position as the initial
> > position.
> [...]
> > Would such a test be fair, or would there be a legitimate objection
> > that Isight is trained on real positions and, with my artificial
> > opening position, there's no guarantee that the positions would be
> > representative of normal play?
> Not only my Isight method is trained on real positions, but humans and
> bots are as well. GNU Backgammon evaluates the position as follows:
>
> 0-ply: No double, take (12.6 % winning chances)
> 1-ply: Double, pass (88.8 % winning chances)
> 2-ply: No double, take (30.7 % winning chances)
> 3-ply: Double, pass (84.0 % winning chances)
> 4-ply: No double take (35.8 % winning chances)
>
> Of course, even a weak (but untruncated) roll-out gives the correct
> result that the player on roll is a slight favorite.
>
> To my surprise, the checker play of the bot was weird in the beginning
> as well (before the game reached more familiar positions), resulting in
> a lot of unnecessary wastage. Together with the bot's premature doubles
> on odd-ply settings (easy takes, and now the cube is on my side) it
> might well be that a human has a good chance to make this a winning
> proposition against the bot.
>
> I will play around a bit, thanks for the idea!
>
> Best regards
>
> Axel

Yes, I think I play better than the bots in this scenario as they make some truly bizarre bear in plays. (at least on regular settings, I'm sure it gets cleaned up a bit on ++ but ++ still suffers from a small search interval so in these positions the best play(s) may not make it through the move filter leaving the computer playing 6/1 with a fiver out of nowhere.

Maybe I'll put it on my fun to do list. It doesn't take very long at all as far as bg projects go.

Stick

On 9/5/2023 4:08 AM, Axel Reichert wrote:
> To my surprise, the checker play of the bot was weird in the beginning
> as well (before the game reached more familiar positions), resulting in
> a lot of unnecessary wastage.

This is a well-known phenomenon.

https://www.bgonline.org/forums/webbbs_config.pl?read=136247
https://www.bgonline.org/forums/webbbs_config.pl?read=136252

---
Tim Chow