AlphaGo vs. AlphaGo with Michael Redmond 9p: Game 33
Vložit
- čas přidán 13. 12. 2019
- Michael Redmond 9p, hosted by the AGA E-Journal’s Chris Garlock, reviews the 33rd game of the amazing AlphaGo vs. AlphaGo selfplay games. The 50 game series was published by Deepmind after AlphaGo's victory over world champion Ke Jie 9p in May 2017.
Check out the sgf file -- with variations/comments -- here: www.usgo.org/news/2019/12/alp...
Produced by Michael Wanek, Andrew Jackson & Allen Moy
Thumbnail image of Rock 'em Sock 'em Robots by Lorie Shaull - Own work, CC BY-SA 4.0, commons.wikimedia.org/w/index... - Hry
So cool to see Redmond again reviewing Alphago games
Thanks! As always, such a pleasure to have Mr. Redmond commenting.
What a great game! Thank you so much for reviewing all these games because without the reviews, I just see so many strange moves that I don't understand, and wouldn't understand which are the real gems. I especially like when Michael admits that he doesn't understand some exchanges either because it shows how godlike these AI are becoming.
Regarding 50:52
With katago b20c256 network (supplied with Lizzie 0.7.2 Engine1):
>400000 playouts move F13 is the "blue move".
With Leela #254 3a258778 network:
>400000 playouts move B8 is the "blue move".
With LeelaMaster_GX78 network:
>400000 playouts move B8 is the "blue move".
Beautiful game and analysis. Thank you. :)
liked, subbed, and commented!!!! Appreciate your good work!!!
I am a simple man. I see a Redmond Review, I like.
And also that kitty!!
*Slight* spoiler if you have not watched the whole game.
Here is my (1kyu) take on why black played the small move at E10 at 7:08 . I think black might have wanted to prevent white playing immediately at B12. Without bE10 in place wB12 would lead to white gouging out black's base with no compensation.
Here is my main line:
wB12 bC12
wB14 bB11
wB13 bB9
wB8 bC8
wA9
Now black can play the ko at A10, but this looks too good for white. If black connects solidly:
... bB10
wC7 bD8
wD7 bE8
wF8
Black needs the solid connection at B10 to make the cut and the clamp work, but this does not have any effect on white's invading stones so white can black this. Black can play B10 instead of attaching at C12, but then white can run out with his invading stone splitting black's group in two.
Excellent game, and explanation: thank you!
I just wonder, at this moment in time (December 15, 2019), me replaying lots of Leela Zero games: my impression is that there is a different feel to these AlphaGo games when compared to the Leela Zero games. These AlphaGo games feel a bit more heroic to me. How do you compare these games from the two engines? Any clue which of the two is stronger right now?
and they're back
The current LZ version (3a258778) does find the geta (52:27) at about 6k simulations on my PC
+
Yes me2, with network from Jan 1st. leela.online-go.com/networks/?C=M&O=D I try to keep updated with the latest weights.
That's interesting because I have an older 15B network, not #157, but a later one trained on the 20B and 40B games, and it was able to immediately find the geta.
Yessss
redmond san at it again
White's C9 pincer is already outside of AlphaGo Teach's database. This is how it would continue the game instead (five variations): online-go.com/demo/view/432630
Everything is almost dead.
Why is Garlock even in this video? All he does is hmm and haw