Zero?

[BCM]

Hi,

There are:
- Leela Zero (game Go)
- Leela Chess Zero
- ChineseChess-AlphaZero
all are free, open-source, and neural network-based, able using CPUs (no need for GPUs anymore).

Isn't there any similar for shogi?

BCM

[Ihor Strumetskyi]

Hi. ☺️

There is AobaZero for shogi. https://github.com/kobanium/aobazero/blob/release/README_en.md

--
You received this message because you are subscribed to the Google Groups "SHOGI-L" group.
To unsubscribe from this group and stop receiving emails from it, send an email to shogi-l+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/shogi-l/ba4ffeb6-afa8-4452-9e82-6c1610837847%40googlegroups.com.

[BCM]

Hi,

thank you very much.

Exactly what I was searching for.

BCM

Am Donnerstag, 4. Juni 2020 18:25:53 UTC+2 schrieb Ihor Strumetskyi:

Hi. ☺️

There is AobaZero for shogi. https://github.com/kobanium/aobazero/blob/release/README_en.md

On Thu, Jun 4, 2020, 19:14 BCM <shog...@arcor.de> wrote:

Hi,

There are:
- Leela Zero (game Go)
- Leela Chess Zero
- ChineseChess-AlphaZero
all are free, open-source, and neural network-based, able using CPUs (no need for GPUs anymore).

Isn't there any similar for shogi?

BCM

--
You received this message because you are subscribed to the Google Groups "SHOGI-L" group.

To unsubscribe from this group and stop receiving emails from it, send an email to sho...@googlegroups.com.

[Shogi Besancon]

Hello everyone, hope everybody is fine,

Do you know if this AI has special features to improve human player skills?

I mean, designing another software with the human defeat as target is not very new, and not very usefull. But if it is designed in view to improve the human skills, it could be interesting. Anything new about that?

Regards,

Gilles

To unsubscribe from this group and stop receiving emails from it, send an email to shogi-l+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/shogi-l/ffd1a27c-6cf2-47c0-8ce1-91d73d22fdf4o%40googlegroups.com.

[h.g.m...@hccnet.nl]

Op Vr, 5 juni, 2020 8:25 am schreef Shogi Besancon:

> Hello everyone, hope everybody is fine,
>
>
> Do you know if this AI has special features to improve human player
> skills?
>
> I mean, designing another software with the human defeat as target is not
> very new, and not very usefull. But if it is designed in view to improve
> the human skills, it could be interesting. Anything new about that?

Problem is that these 'Zero' engines are not really designed at all. They
just learned to play the game themselves, and even the programmers have no
clue as how they eventually manage to do it. It is not only that they do
not teach humans to play better Shogi, they also do not teach programmers
how to make better Shogi engines...

>>>> <https://groups.google.com/d/msgid/shogi-l/ba4ffeb6-afa8-4452-9e82-
>>>> 6c1610837847%40googlegroups.com?utm_medium=email&utm_source=footer>
>>>> .

>>>>
>>>>
>>> --
>>>
>> You received this message because you are subscribed to the Google
>> Groups
>> "SHOGI-L" group.
>> To unsubscribe from this group and stop receiving emails from it, send
>> an email to shogi-l+u...@googlegroups.com. To view this discussion
>> on the web visit
>> https://groups.google.com/d/msgid/shogi-l/ffd1a27c-6cf2-47c0-8ce1-91d73
>> d22fdf4o%40googlegroups.com

>> <https://groups.google.com/d/msgid/shogi-l/ffd1a27c-6cf2-47c0-8ce1-91d7
>> 3d22fdf4o%40googlegroups.com?utm_medium=email&utm_source=footer>
>> .

>>
>>
>
> --
> You received this message because you are subscribed to the Google Groups
> "SHOGI-L" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to shogi-l+u...@googlegroups.com. To view this discussion on
> the web visit

> https://groups.google.com/d/msgid/shogi-l/CAFGzfKEOCaB4%2Bn7R3LSbzWJBgQTd
> GH%2BcuqGSpo9oCCxKGe_v6g%40mail.gmail.com.
>
>

[captbirdseye]

On Friday, 5 June 2020 09:18:51 UTC+1, H.G.Muller wrote:

Problem is that these 'Zero' engines are not really designed at all. They
just learned to play the game themselves, and even the programmers have no
clue as how they eventually manage to do it. It is not only that they do
not teach humans to play better Shogi, they also do not teach programmers
how to make better Shogi engines...

Ah! A that's bit scary but it made me smile...😎 

How very true - my current programming project (music program, unrelated to Shogi) has a 

couple of very spotty blocks of code where I'm convinced that the code should not work - but 

it does, and I can't see why. This is far worse than the code not working and being unable to

see why... 😞

[BCM]

This 'Zero' engines are not good for teaching, but still we can have some benefit of them and also get interesting things.

1) for example read this: https://arxiv.org/pdf/1712.01815.pdf (page 6)
it looks like this AI is simulating all the centuries of human doing (in a few hours).
e.g. the Caro-Kann opening: after 2 hours learning, the AI decide it's a good opening and used it frequently. But after 6 hours learning, it decides no longer use it.
Maybe a chess-player who have a good and comfrtable history chess-db is reading this: it would be very interesting (for me) to have this 12 graphics also for human-only-games.
So e.g., how often Caro-Kann was played for each year in history by humans?
Did this AI and H(uman)I have the same development?

And what about shogi?
Such learning-development would be also very interesting there.
Did the AI also learned e.g. Anaguma castle by itself or didn't play it at all?
Or how about ranging-/static-rook usage during learning phase?
Following the change of playing-style would be very interesting

2) see the book "Rethinking Opening Strategy: Alphago's Impact on Pro Play"  by Yuan Zhou
Indeed human can learn from it "The AI program AlphaGo Zero has introduced several new ideas about how to play ..."
And how about 8th move in this chess game https://www.youtube.com/watch?v=A-vNq61KfLs?
Any similar in shogi?

[Ian Osgood]

I would also be interested in seeing the Zero-methodology applied to another national chess variant: Thai Makruk. It has more strict endgame counting rules, which if encoded correctly might overcome the stylistic complaint of the NN-based chess engines, that they play too lax when a) far ahead and b) very even due to the one-size-fits-all 50-move rule.

> Did the AI also learned e.g. Anaguma castle by itself or didn't play it at all?

From what I remember, it did not discover the traditional castles like Anaguma, Yagura, and Mino.

> Or how about ranging-/static-rook usage during learning phase?

The papers I saw did have interesting graphs in that area, and showed a vast preponderance towards double static rook. I wonder if that has influenced Sota Fujii's choice of opening.

[h.g.m...@hccnet.nl]

Op zo., aug. 23, 2020 om 19:44, Ian Osgood <ia...@quirkster.com> schreef:

I would also be interested in seeing the Zero-methodology applied to another national chess variant: Thai Makruk. It has more strict endgame counting rules, which if encoded correctly might overcome the stylistic complaint of the NN-based chess engines, that they play too lax when a) far ahead and b) very even due to the one-size-fits-all 50-move rule.

I don't expect Makruk to be very different. The counting rules don't seem to affect the game all that much, as they only spring into action when there are no Pawns. So the easiest way to deal with them is to never take the opponent's last Pawn, unless you already see the mate. This is probably what the AI would learn, and it would be enough.

The lax play is not intrinsic to NN-based engines, but merely a consequence of wrong training. These things are trained for not being able to convert winning positions; as long as they are still able to make progress in certainly-won positions, their evaluation is not considered perfect, and will be continued until they cannot see the difference between a certain quick win and a certain slow win anymore.

With these NNs you pretty much get what you ask for.

 

> Did the AI also learned e.g. Anaguma castle by itself or didn't play it at all?
From what I remember, it did not discover the traditional castles like Anaguma, Yagura, and Mino.

> Or how about ranging-/static-rook usage during learning phase?

The papers I saw did have interesting graphs in that area, and showed a vast preponderance towards double static rook. I wonder if that has influenced Sota Fujii's choice of opening.

On Monday, June 8, 2020 at 7:55:13 AM UTC-7 BCM wrote:

This 'Zero' engines are not good for teaching, but still we can have some benefit of them and also get interesting things.

1) for example read this: https://arxiv.org/pdf/1712.01815.pdf (page 6)
it looks like this AI is simulating all the centuries of human doing (in a few hours).
e.g. the Caro-Kann opening: after 2 hours learning, the AI decide it's a good opening and used it frequently. But after 6 hours learning, it decides no longer use it.
Maybe a chess-player who have a good and comfrtable history chess-db is reading this: it would be very interesting (for me) to have this 12 graphics also for human-only-games.
So e.g., how often Caro-Kann was played for each year in history by humans?
Did this AI and H(uman)I have the same development?

And what about shogi?
Such learning-development would be also very interesting there.
Did the AI also learned e.g. Anaguma castle by itself or didn't play it at all?
Or how about ranging-/static-rook usage during learning phase?
Following the change of playing-style would be very interesting

2) see the book "Rethinking Opening Strategy: Alphago's Impact on Pro Play"  by Yuan Zhou
Indeed human can learn from it "The AI program AlphaGo Zero has introduced several new ideas about how to play ..."
And how about 8th move in this chess game https://www.youtube.com/watch?v=A-vNq61KfLs?
Any similar in shogi?

--
You received this message because you are subscribed to the Google Groups "SHOGI-L" group.
To unsubscribe from this group and stop receiving emails from it, send an email to shogi-l+u...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/shogi-l/f69224e0-26f1-4a1e-a548-f6b5a37aff4bn%40googlegroups.com.