Lee Sedol 9p-Yamashita Keigo 9p

[Elom]

I've wondered for quite a while on what bot's may think of this game.

I put the commentary in below, from which it might be clear why I'm curious (black leads by a lot early on):

[Tryss]

According to LZ, black is not ahead by "a lot". At move 91, black winrate is only 63%. After the sucessful invasion on the left side, white is ahead

[eyecatcher]

i read in the "the style of Lee Sedol" book, that it is typical for his games/style... being behind in the opening/beginning and than winning/catching up anyway by initiating complex middle game fighting

[nasdaq]

thanks for sharing!, all those ko fights so early make me nervous
lee is tenacious, not giving up

[Uberdude]

I could do a LZ review like forum/viewtopic.php?f=13&t=15924 if people are interested.

[dust]

Yamashita's view on the game might be quite interesting.

It looks like one of those terrible setbacks that take you about a month to recover from and regain your self-confidence.

[baduk]

alot from pros perspective is a difference of 5-7 points:)

[Elom]

thanks for sharing!, all those ko fights so early make me nervous
lee is tenacious, not giving up

Your welcome— commentary from An Youngill 8p, who was prolific with 105 reviews. It's a good thing I collected them all before the site went down.

i read in the "the style of Lee Sedol" book, that it is typical for his games/style... being behind in the opening/beginning and than winning/catching up anyway by initiating complex middle game fighting

Yes, one might guess that studying his games simply for his spirit might do wonders alone.

I could do a LZ review like forum/viewtopic.php?f=13&t=15924 if people are interested.

Yes please.

[Uberdude]

The beginnings of an in-depth LZ review plus joseki diversions (2+ hours on the first 15 moves!):



Update: done to move 48.

[Elom]

Thanks! This is brilliant .

[Uberdude]

Here's LZ 40 block's winrate graph for the game on 10k playouts (black win at top).

Lee Yamashita winrate annotated.PNG (475.38 KiB) Viewed 20101 times

Also I updated the sgf with commentary to move 48.

[ez4u]

Would you be willing to run the same analysis with the current 20-block to compare the win-rate graphs?

[Uberdude]

Sure, here's LZ #162 (20 blocks) on 10k playouts. I suppose I should have used GRP to export the numbers so I can graph them together . Shapes are pretty similar, I think most significant difference is when Yamashita makes big mistake in middlegame and white's winning and then Lee makes smaller mistake back 40b thinks white is still winning plenty whereas 20b thinks it bounced back to a black win for some moves.

Lee Yamashita LZ162 winrate.PNG (469.71 KiB) Viewed 20053 times

[kvasir]

We were discussing KataGo's veiw on games from Lee Sedol's commented games volume 1 over here when this game was mentioned.

I ran analysis with Katrain using kata1-b40c256-s10336005120-d2519775087.bin.gz which just happened to be one of the better weights when I downloaded weights and also happened to be b40 instead of a larger network.

I ran 2.5k, 10k and 100k playouts (which took awhile because of all the comments) and the graphs follow.

2500graph.png (54.45 KiB) Viewed 14165 times

10000graph.png (53.81 KiB) Viewed 14165 times

100000graph.png (54.1 KiB) Viewed 14165 times

There is not much difference in the overall picture except that the score and win rate graphs become slightly smoother and the win rate does tend more toward the extremes. There are probably some instances when the score estimate is shifted horizontally, it is a bit hard to tell and I am not going to compare the individual runs.


As for the game overall:

The game is close to error free considering that the players didn't have programs like KataGo to refresh on the more sublime aspects of the game. There are only 10 mistakes that are greater than 1.667 points and many of the larger ones are very hard to understand.

The sets of point losses >=1 point are:
Black: 4.0 3.6 1.9 1.7 1.6 1.6 1.5 1.4 1.4 1.4 1.4 1.2 1.2 1.2 1.1 1.0 Total: 27.2
White: 3.5 2.9 2.7 2.5 2.0 1.7 1.7 1.6 1.5 1.5 1.5 1.4 1.2 1.1 Total: 26.8
So in terms of mistakes the players are only ~0.4 points apart.

To me it is a clear case of "who makes the last mistake loses", I am not sure there is much more to it because this game is very well played and never really far away from even.

As for the review:

It mostly misses the largest errors, I think because these are very hard to call even though KataGo finds them easily. It also says black is ahead for the first part of the game, even when KataGo is claiming a 0.5 game. It praises black with the words "Yamashita plays very well when he takes the lead in the game." At that point black has lost the lead in eyes of KataGo and it is W+0.3. It is tough to compete with KataGo's positional judgment.

Examples of hard to call errors:

Moves 54 - 54. Both are ~3.5 point mistakes. KataGo is going to let white push through the bamboo joint.
My take: This is not Human play.

Moves 106 - 107. OK, white might be timing 106 badly or possibly trying to change the flow. It is a 2.9 point error. KataGo is just going to tenuki and leave 8 stones for a ko that is hard to explain. Black 107 is a 4.0 point error.
My take: 4.0 - 2.9 is 1.1 so black is not making a mistake, he is avoiding complications.

Effectively we could remove some of the largest errors spotted by KataGo from consideration. It seems unrealistic to try to copy KataGo in these cases.


There are a two cases when KataGo completely contradicts the comments. I think the lesson is that if there is a move that you want to make then you really need to try to make it work; that is what KataGo seems to do. KataGo seems to work very hard to make the obvious move work when it "obviously" doesn't work. The two cases:

Move 61: "B really wants to answer at A, but there's no following ko threat for B."
KataGo: "Play A and G18 is my threat."
My take: KataGo is right. G18 doesn't look like a beefy threat but black survives and has that move in the center that we really want.

Move 81: "W can't answer this as W doesn't have any big ko threats."
KataGo: "Just answer."
My take: The commentary is correct. Black becomes too thick in the center if you answer and white can't handle this game then.


I can't say much about the game myself, even having all the aid from KataGo, it is just too complicated and too well played. I think the commentary is good and my feeling is that it is not meant to be a very objective analysis that dives deep but more as a guide to help people follow along. It is educational and fun, even if in my opinion (now biased by looking at the game in Katrain) it exaggerates black's advantage in the early game.

I might consider adding diagrams if there is demand, they are usually tedious to make. It could also be interesting to go into how black should have handled the invasion on move 119.

[Elom0]

We were discussing KataGo's veiw on games from Lee Sedol's commented games volume 1 over here when this game was mentioned.

I ran analysis with Katrain using kata1-b40c256-s10336005120-d2519775087.bin.gz which just happened to be one of the better weights when I downloaded weights and also happened to be b40 instead of a larger network.

I ran 2.5k, 10k and 100k playouts (which took awhile because of all the comments) and the graphs follow.

2500graph.png

10000graph.png

100000graph.png

There is not much difference in the overall picture except that the score and win rate graphs become slightly smoother and the win rate does tend more toward the extremes. There are probably some instances when the score estimate is shifted horizontally, it is a bit hard to tell and I am not going to compare the individual runs.


As for the game overall:

The game is close to error free considering that the players didn't have programs like KataGo to refresh on the more sublime aspects of the game. There are only 10 mistakes that are greater than 1.667 points and many of the larger ones are very hard to understand.

The sets of point losses >=1 point are:
Black: 4.0 3.6 1.9 1.7 1.6 1.6 1.5 1.4 1.4 1.4 1.4 1.2 1.2 1.2 1.1 1.0 Total: 27.2
White: 3.5 2.9 2.7 2.5 2.0 1.7 1.7 1.6 1.5 1.5 1.5 1.4 1.2 1.1 Total: 26.8
So in terms of mistakes the players are only ~0.4 points apart.

To me it is a clear case of "who makes the last mistake loses", I am not sure there is much more to it because this game is very well played and never really far away from even.

As for the review:

It mostly misses the largest errors, I think because these are very hard to call even though KataGo finds them easily. It also says black is ahead for the first part of the game, even when KataGo is claiming a 0.5 game. It praises black with the words "Yamashita plays very well when he takes the lead in the game." At that point black has lost the lead in eyes of KataGo and it is W+0.3. It is tough to compete with KataGo's positional judgment.

Examples of hard to call errors:

Moves 54 - 54. Both are ~3.5 point mistakes. KataGo is going to let white push through the bamboo joint.
My take: This is not Human play.

Moves 106 - 107. OK, white might be timing 106 badly or possibly trying to change the flow. It is a 2.9 point error. KataGo is just going to tenuki and leave 8 stones for a ko that is hard to explain. Black 107 is a 4.0 point error.
My take: 4.0 - 2.9 is 1.1 so black is not making a mistake, he is avoiding complications.

Effectively we could remove some of the largest errors spotted by KataGo from consideration. It seems unrealistic to try to copy KataGo in these cases.


There are a two cases when KataGo completely contradicts the comments. I think the lesson is that if there is a move that you want to make then you really need to try to make it work; that is what KataGo seems to do. KataGo seems to work very hard to make the obvious move work when it "obviously" doesn't work. The two cases:

Move 61: "B really wants to answer at A, but there's no following ko threat for B."
KataGo: "Play A and G18 is my threat."
My take: KataGo is right. G18 doesn't look like a beefy threat but black survives and has that move in the center that we really want.

Move 81: "W can't answer this as W doesn't have any big ko threats."
KataGo: "Just answer."
My take: The commentary is correct. Black becomes too thick in the center if you answer and white can't handle this game then.


I can't say much about the game myself, even having all the aid from KataGo, it is just too complicated and too well played. I think the commentary is good and my feeling is that it is not meant to be a very objective analysis that dives deep but more as a guide to help people follow along. It is educational and fun, even if in my opinion (now biased by looking at the game in Katrain) it exaggerates black's advantage in the early game.

I might consider adding diagrams if there is demand, they are usually tedious to make. It could also be interesting to go into how black should have handled the invasion on move 119.

Wow this is amazing !