New networks Leela Zero, a Sensible Artificial Intelligence

[Bill Spight]

Thanks for SAI's variations.
Not solved, I am afraid. Here is an SGF file.

[And]

CS Zero probably sees some trick :
C19 - 57.6% 85178026
A17 - 56.5% 155301
H16 - 55% 122714
S14 - 56.1% 85306
A7 - 54.7% 65181

playing on dan 9 chooses a move H2

[georgenan]

Sorry for my unclear description.

It's is a test for professionals in endgame stage and it's designed by Berlekamp https://math.berkeley.edu/~berlek/index.html

It's Japanese Rule and its goal is to give the right sequence with no concession.

White first. When komi>2, it's easy to win for bots.

Since Katago is less-consolidation with territory-counting and SAI is less-consolidation with komi-estimation,
But SAI's 9b will give wrong answer in live-or-dead situation.
In this game, SAI always makes wrong judgement on top left 10 stones.

My SAI parameters:

Code: Select all

-g -b 0  --noponder -p 0 -v 0 -r 3 -m 0 --komi 0 --nrsymm --lambda 0.1 --mu 0.5 -t 6 --batchsize 5 -j

SAI performs more accurate than KataGo when komi = 0.

I'm using weight file SAI238.

Using --lambda 0.5 SAI may try aggressive moves. Since --lambda 0.1 is used in selfplays I think it could be a proper choice.
Using --mu less than 0.2 will give H2 as other strong bots.

[georgenan]

S14 is the best move by designer Prof. Berlekamp and Chang Hao gave the right answer in news report.
But the mathematical analysis method hadn't accepted by go players.
I read this on a Chinese articlehttps://zhuanlan.zhihu.com/p/33860825 without the sequence after S14.

[Bill Spight]

It's is a test for professionals in endgame stage and it's designed by Berlekamp https://math.berkeley.edu/~berlek/index.html

This problem is based upon Theorem 8 in the book, which David Wolfe solved for his Ph.D. thesis.

It's Japanese Rule and its goal is to give the right sequence with no concession.

The problems in the book lead to the same answer under both Japanese and Chinese rules, as with correct play White wins by 1 pt. with 0 komi. Generally there is no single correct sequence, and not always only one correct first play. Amateurs who have understood the book can solve these problems in a minute or two.

In this game, SAI always makes wrong judgement on top left 10 stones.

I take it that you are referring to the 10 White stones that eventually will be put in atari. One important aspect of this position is that with correct play that atari will not happen before the dame stage. It is unlikely that a bot will have learned those conditions through self play. That fact makes plays gain less than 1 pt. by territory scoring and less than 2 pts. by area scoring. IMHO the book does not explain the conditions under which that is the case clearly enough for practical go players who are not mathematicians. Oven 10 years ago I wrote up a small article which I hope does a better job of explaining Wolfe's results. I would like to get it published in a top weiqi magazine.

[Bill Spight]

S14 is the best move by designer Prof. Berlekamp and Chang Hao gave the right answer in news report.

David Wolfe may be the composer of that problem, as it is based upon his theorem. Anyway, you can show that White S-14 is the best firs move, but four other first moves produce a White win with correct play.

But the mathematical analysis method hadn't accepted by go players.

This problem may be analyzed using difference games, which Berlekamp and Wolfe explain in their book, and by the method of multiples, which I deveolped to illustrate the mean value theorem. I have written about both methods online for many years. Instead of the method of multiples it may be analyzed by thermography, which also appears in the book, or by other methods which I discovered on my own, and I am sure other people have, as well. See Robert Jasiek's books.

But, yes, sadly, the methods in the book have not been adopted by pros, even though the book came out in 1994, and other papers on these topics have been published since. I ran into unexpected opposition a few years ago in getting my article about getting the last play published in the magazine Myosu. The opposition came not from a pro, but from a strong amateur on their editorial staff. {shrug}

[And]

networks of 12x256 appeared in test matches, soon there will probably be a change http://sai.unich.it

[Javaness2]

I am rather late to the party, so forgive me if my question sounds a bit rude, but...

What is different about this HNG variant of LeelaZero, why should we be interested in it?

[And]

https://github.com/sai-dev/sai/wiki/Features

[And]

Probably today there will be a transition to 12x256

[And]

"Hello! We are moving to 12x256 with the next promotion. Stay tuned!"
https://github.com/sai-dev/sai/issues/3 ... -586396400

[lichigo]

Hello, I'm wondering what is the difference between 40x256 and 12x256?

[counting]

Hello, I'm wondering what is the difference between 40x256 and 12x256?

Obviously 40x256 is 10240 and 12x256 is 3072, so 40x256 has more power

Kidding aside, how seriously do you want your answer?

If you only want a graphic representation of the structure, LZ's structure is similar to AlphaGo Zero
https://i.imgur.com/2e4iyAy.jpg

If you want to know what each technical term means, there is a technical explanation for Leela Chess Zero, and the terminologies and technologies used are the same.
https://github.com/LeelaChessZero/lc0/w ... Chess-Zero

If you really want to know the details you can go to their github page and read all the AlphaGo papers by Deepmind
https://github.com/leela-zero/leela-zer ... /README.md

And as the number suggested, more blocks it has, more complex patterns it can remember and has a better instinct to find the next move.

[splee99]

Don't forget the capability to play handicap games. Although the 40 block network is powerful, it basically has not much idea how to play handicap games.

[And]

SAI already in some positions smarter than the ELFv2! for example, understands the ladder!
ELFv2 - SAI (#285 aa09e39f), GF610, 5sec/move:


SAI's next move is p8!
I changed the color of the engines in this position, the ELF continued to the ladder and after a few moves gave up!