Life In 19x19
http://lifein19x19.com/

Unusual case of efficiency
http://lifein19x19.com/viewtopic.php?f=13&t=15545
Page 2 of 2

Author:  Gomoto [ Wed Mar 28, 2018 6:13 pm ]
Post subject:  Re: Unusual case of efficiency

Variation a:
Attachment:
john1.jpg
john1.jpg [ 305.3 KiB | Viewed 1545 times ]

Author:  Gomoto [ Wed Mar 28, 2018 6:14 pm ]
Post subject:  Re: Unusual case of efficiency

Variation b:
Attachment:
john2.jpg
john2.jpg [ 319.82 KiB | Viewed 1545 times ]

Author:  Gomoto [ Wed Mar 28, 2018 6:24 pm ]
Post subject:  Re: Unusual case of efficiency

Leela Zero won against Ha-jin Lee. It is definitly pro strength now. I am very interested to see further test games against top pros or top AIs.

Zen 7 struggles to analyze complicated fights between the likes of Ke Jie and Lee Sedol. I have yet to try analyzing such games with Leela Zero.

Author:  sorin [ Wed Mar 28, 2018 6:27 pm ]
Post subject:  Re: Unusual case of efficiency

Kirby wrote:
When Leela gives a win rate for black (e.g. 59%), is that self play or just evaluation?


It is both: basically the same MCTS process that AlphaGo uses to decide the moves.

As for how strong it is (your earlier question): I think LeelaZero should be somewhere between AlphaGo-FanHui and AlphaGo-Lee I believe.

Not sure if we should believe it or not when it disagrees with top pros, but in both pro games in this thread it has indeed very strong opinions against the pro choices...

Author:  Gomoto [ Wed Mar 28, 2018 6:36 pm ]
Post subject:  Re: Unusual case of efficiency

I do not think it is a question of believe, I always vote for comparing variations :)

There is probably no absolute truth in Go (thanks Ko). But even if there is only one right way to play the AIs still have a long way to go.

I do not take the winning percentages as absolute truth (Competing AI´s differ wildly in their OPINIONS by the way.)
I interpret the percentages as the opinions of competing very strong, and therefore respectable, players.

[Edit:] (Dont get me wrong I respect players of any strength, here I am talking only about respectable in the sense: you have to check out the variations of these strong AIs to get to your own opinion.)

Author:  Kirby [ Wed Mar 28, 2018 7:32 pm ]
Post subject:  Re: Unusual case of efficiency

Gomoto wrote:
I do not think it is a question of believe, I always vote for comparing variations :)

There is probably no absolute truth in Go (thanks Ko). But even if there is only one right way to play the AIs still have a long way to go.

I do not take the winning percentages as absolute truth ...


Maybe I misunderstood. I didn't get that impression from earlier discussions. e.g.:

Gomoto wrote:
White to play. There is only one good move for white. Are you the first to find the tesuji?
...
Thanks for participating, the solution and only move is the tesuji at P6


Anyway, comparing variations is fine, but evaluating variations is still difficult. So I think there still has to be some element of "belief" when you trust what a bot says. You can see the variation it gives, and maybe it gives you a new idea. But you have to come up with the reasoning as to why a given variation is good yourself.

sorin wrote:
It is both: basically the same MCTS process that AlphaGo uses to decide the moves.

As for how strong it is (your earlier question): I think LeelaZero should be somewhere between AlphaGo-FanHui and AlphaGo-Lee I believe.

Not sure if we should believe it or not when it disagrees with top pros, but in both pro games in this thread it has indeed very strong opinions against the pro choices..


Thanks for explaining about Leela's strength. And you bring up a good point about the strong opinions the bot has.

If I get time, I might play around with it. I'm assuming this is the program: https://github.com/gcp/leela-zero

Author:  Bill Spight [ Wed Mar 28, 2018 8:17 pm ]
Post subject:  Re: Unusual case of efficiency

Kirby wrote:
When Leela gives a win rate for black (e.g. 59%), is that self play or just evaluation?


IIUC, it's just evaluation, unless otherwise specified. Strictly speaking, the evaluations of neural network bots are not actually win rates, unlike those of pure MCTS bots. And the MCTS win rates are for quasi-random playouts.

Author:  sorin [ Wed Mar 28, 2018 8:32 pm ]
Post subject:  Re: Unusual case of efficiency

Kirby wrote:
sorin wrote:
It is both: basically the same MCTS process that AlphaGo uses to decide the moves.

As for how strong it is (your earlier question): I think LeelaZero should be somewhere between AlphaGo-FanHui and AlphaGo-Lee I believe.

Not sure if we should believe it or not when it disagrees with top pros, but in both pro games in this thread it has indeed very strong opinions against the pro choices..


Thanks for explaining about Leela's strength. And you bring up a good point about the strong opinions the bot has.

If I get time, I might play around with it. I'm assuming this is the program: https://github.com/gcp/leela-zero


Right, that is the source code. For Windows, I download the latest precompiled version listed here: http://zero.sjeng.org/ and the network weights file from the same page, and I use Sabaki to interact with leelaz.exe via GTP.

Author:  Calvin Clark [ Wed Mar 28, 2018 9:31 pm ]
Post subject:  Re: Unusual case of efficiency

I like ez4u's analysis. I can't add to it.

But I will say that it's interesting how unlikely this joseki would be to show up in games involving top bots. If you're a player copying AlphaGo cargo-cult style, then:

You'd probably prefer :b1: to any loose pincer like 'a'.

Click Here To Show Diagram Code
[go]$$B
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . . O . . . a
$$ | . . X , . . . . .
$$ | . . . . . . . . .
$$ | . . . 1 . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]


If pincered, you'll have an uncontrollable urge to press at :w1: rather than playing some settling move at 'a' or 'b'.

Click Here To Show Diagram Code
[go]$$W
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . b . O . . . X
$$ | . . X a . . . . .
$$ | . . . 1 . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]


Lastly, somehow I found a pattern in AlphaGo Teach. You have to play a lot of human, non-AlphaGo moves to get here, but finally, AlphaGo would extend at :w1: and not 'a'. With that move, it gives a black winning percentage of 40.7%.

Click Here To Show Diagram Code
[go]$$W
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . X X O O . 1 . . . . . . . . . . . . |
$$ | . . . X . . a . . , . . . . . X . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


I won't go so far as to say forget this joseki, but...well, it's up to you. :)

Author:  Bill Spight [ Wed Mar 28, 2018 11:13 pm ]
Post subject:  Re: Unusual case of efficiency

Calvin Clark wrote:
I won't go so far as to say forget this joseki, but...well, it's up to you. :)


Great book title: Forget This Joseki :mrgreen:

Author:  Gomoto [ Thu Mar 29, 2018 4:07 am ]
Post subject:  Re: Unusual case of efficiency

Lets make a thread ;-)





and by the way
cargo cult :lol: :tmbup:

Page 2 of 2 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/