Life In 19x19

xela wrote: If you have a text editor that can do search and replace using regular expressions, then you can probably get rid of all the LZ[...] tags by replacing "\[.*?\n.*\];" with ";". If this doesn't work for you, I'm happy to discuss other options either here or by PM. I'm keen to see your analysis

I got this far in understanding the problem but haven't yet found an editor with regex capability. I was down to using python but then sloth got the better of me. If I send you the file, would you be so kind ... ?

You can open the file in Lizzie, select "clear evaluations" or some option like that in the menu, then save it. It will still have all the analysis tags there, but they'll be wiped to just a dummy "0 visits, no evaluation" tag, which takes much less space.

lightvector wrote:You can open the file in Lizzie, select "clear evaluations" or some option like that in the menu, then save it. It will still have all the analysis tags there, but they'll be wiped to just a dummy "0 visits, no evaluation" tag, which takes much less space.

In the "Analyze" menu, there's an item that says "clear analysis", but for me it only removes the analysis info from the current move, not from the entire SGF file. Am I missing something here?

Knotwilg wrote: I got this far in understanding the problem but haven't yet found an editor with regex capability. I was down to using python but then sloth got the better of me. If I send you the file, would you be so kind ... ?

No problem. Email is probably easiest.

xela wrote:

lightvector wrote:You can open the file in Lizzie, select "clear evaluations" or some option like that in the menu, then save it. It will still have all the analysis tags there, but they'll be wiped to just a dummy "0 visits, no evaluation" tag, which takes much less space.

In the "Analyze" menu, there's an item that says "clear analysis", but for me it only removes the analysis info from the current move, not from the entire SGF file. Am I missing something here?

Eeew. Okay then. Regex it is.

Thanks to xela, here's the sgf:

Many thanks to Knotwilg (and xela) for showing us this interesting game.

Quick comment on . Looking at the Elf commentary, I expect that the suspicious drop in White's winrate estimate is the result of too few rollouts, which makes the winrate estimate unreliable. For instance, Elf recommends the jump attachment at F-09, giving it 107k rollouts, while Otake's play gets only 2.1k rollouts. Elf estimates the White winrate after as 68%, but after , which Elf thinks is correct, it estimates White's winrate as 71½%, with 31.6k rollouts. Do we really think that was a mistake that boosted White's chances by 3½%? Or do we think that it provides a better winrate estimate with 15 times as many rollouts?

Similar remarks apply to the winrate estimate after . KataGo estimates White's winrate at 25%. But on how many rollouts?

Elf estimates White's winrate after at 23½%, but with only 678 rollouts. It then agrees with Kajiwara's play, , and then estimates White's winrate at 34½% with 30,641 rollouts. Do we think that despite being both Kajiwara's and ELf's choice, was a mistake that increased White's chances by 11%? Or is 34½% a better estimate of White's chances, based upon more than 45 times as many rollouts?

Like KataGo, Elf missed , giving it only 1 rollout. It didn't entirely miss it, it just wanted to play the sequence W D-11 - B C-11, W B-10, B B-09, first, thereby sacrificing an extra stone on the left side unnecessarily. After Otake's it estimates White's winrate at 49%, while it estimates White's winrate after its top choice for White of D-11 at 35½% (with 59.8k rollouts). IOW, it rates its top choice as losing 13% to Otake's play. (BTW, its estimate of 49% for that play is not based upon only 1 rollout. It inherits its estimate from , which has 14.2k rollouts.)

OC, Elf's blindspot makes its winrate estimates for and earlier plays incorrect. But that does not justify the winrate estimate for on only 678 rollouts. In fact, the earlier estimates should be revised upwards for White.

There is a remarkable discrepancy between the winrate estimates of KataGo and the Elf commentary. As we know, Elf tends to favor Black by comparison with other bots, at least at first, and its evaluations tend to be more extreme than other bots'. However, at move 181, where a ko fight breaks out, KataGo gives Black a 61% winrate estimate, at that same point, Elf gives Black only a 24½% winrate estimate (with 47.5k rollouts), a discrepancy of 36½%, not counting the original difference at move 1.

Thanks, I really enjoyed going through this game! I too struggled to understand the status of the white groups on the right at various times. And was a neat tesuji. But for me the most surprising move of the game was .

Click Here To Show Diagram Code

[go]$$Wc
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . O . . . . . . . . . . . |
$$ | . . . O . O . . . . . . . . . . O . . |
$$ | . . O . . O X . X . . . . X . X X . . |
$$ | . . O X . . X . . . . X . . . . X . . |
$$ | . . X . . . . . . . . b . . . O O O . |
$$ | . . . X . . . . . . . . X . X . . . . |
$$ | . . . . . . . . W W W W X . . . O . . |
$$ | . . . . . . . . W X X X O O O O . X . |
$$ | . . . . . . . . . . . . X . X . . O . |
$$ | . . X . . . . . W . X . X . . . . O . |
$$ | . . . . . . . . . . . . . . X X X X . |
$$ | . . O . X . . . . . . . O O . O . . . |
$$ | . . . a . . . . . . . . X O . . . O . |
$$ | . . . . . . . . . . . X . X . O . . . |
$$ | . . . O . . . . . X . . X . . O . O . |
$$ | . . . . . . . . . . . . X . X X O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I thought a move to strengthen the centre group was pretty much essential at this point. But no, Otake fearlessly played a! KataGo does indeed prefer a move in the centre (and has several suggestions, not a single clear favourite), but thinks that the game move is still playable (white's lead drops to 2 points). In many variations, white has b as a saving move.

That is a "+11% move" for my KataGo 1.3 setup. That is, KataGo doesn't find the move, but when I put it on the board, the winrate is 11% above KataGo's first choice. Also, is +8% (KataGo just looks at the clamp instead), and is +4%. With these apparent blind spots, I agree that we can't trust KataGo's evaluations of other moves round about this part of the game. In particular, I don't think we can conclude that is actually a blunder.

In hindsight, after seeing I wondered if was a mistake. But any alternatives appear to be even worse.

Earlier, I'm pleased to find that I seem to have learned something from Genjo-Chitoku: I was able to recognise as creating momentum for (although KataGo isn't overly impressed by and prefers a pincer for once).

Bill Spight wrote:There is a remarkable discrepancy between the winrate estimates of KataGo and the Elf commentary. As we know, Elf tends to favor Black by comparison with other bots, at least at first, and its evaluations tend to be more extreme than other bots'. However, at move 181, where a ko fight breaks out, KataGo gives Black a 61% winrate estimate, at that same point, Elf gives Black only a 24½% winrate estimate (with 47.5k rollouts), a discrepancy of 36½%, not counting the original difference at move 1. :o

Not at all remarkable. Remember that KataGo can work with the correct komi in these old games, but ELF can't. On 10k playouts, I'm seeing Black at +61% or +0.8 points in KataGo with komi=5, or +47%, +0.3 points if I edit the SGF file to the correct komi of 5.5. With the position so finely balanced (and late enough in the game that score estimation should be pretty accurate), it's no surprise that an extra two points of komi will change the winrate a lot.

xela wrote:

Bill Spight wrote:There is a remarkable discrepancy between the winrate estimates of KataGo and the Elf commentary. As we know, Elf tends to favor Black by comparison with other bots, at least at first, and its evaluations tend to be more extreme than other bots'. However, at move 181, where a ko fight breaks out, KataGo gives Black a 61% winrate estimate, at that same point, Elf gives Black only a 24½% winrate estimate (with 47.5k rollouts), a discrepancy of 36½%, not counting the original difference at move 1.

Not at all remarkable. Remember that KataGo can work with the correct komi in these old games, but ELF can't. On 10k playouts, I'm seeing Black at +61% or +0.8 points in KataGo with komi=5, or +47%, +0.3 points if I edit the SGF file to the correct komi of 5.5. With the position so finely balanced (and late enough in the game that score estimation should be pretty accurate), it's no surprise that an extra two points of komi will change the winrate a lot.

Oh, right. I was not figuring in the different komi settings. Viva KataGo!

xela wrote: Earlier, I'm pleased to find that I seem to have learned something from Genjo-Chitoku: I was able to recognise as creating momentum for (although KataGo isn't overly impressed by and prefers a pincer for once).

Remember that the value of pushing or jumping through an opening is as a rule better the smaller the opening is. Jumping through a 5 space opening is not a big deal. In fact, for Elf prefers the jaw jab at R-10 by 3½% to the jump. Curiously, Elf prefers the jump for , but only by 1% over the kick.

xela wrote:But for me the most surprising move of the game was .

$$Wc
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . O . . . . . . . . . . . |
$$ | . . . O . O . . . . . . . . . . O . . |
$$ | . . O . . O X . X . . . . X . X X . . |
$$ | . . O X . . X . . . . X . . . . X . . |
$$ | . . X . . . . . . . . b . . . O O O . |
$$ | . . . X . . . . . . . . X . X . . . . |
$$ | . . . . . . . . W W W W X . . . O . . |
$$ | . . . . . . . . W X X X O O O O . X . |
$$ | . . . . . . . . . . . . X . X . . O . |
$$ | . . X . . . . . W . X . X . . . . O . |
$$ | . . . . . . . . . . . . . . X X X X . |
$$ | . . O . X . . . . . . . O O . O . . . |
$$ | . . . a . . . . . . . . X O . . . O . |
$$ | . . . . . . . . . . . X . X . O . . . |
$$ | . . . O . . . . . X . . X . . O . O . |
$$ | . . . . . . . . . . . . X . X X O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+

Click Here To Show Diagram Code

[go]$$Wc
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . O . . . . . . . . . . . |
$$ | . . . O . O . . . . . . . . . . O . . |
$$ | . . O . . O X . X . . . . X . X X . . |
$$ | . . O X . . X . . . . X . . . . X . . |
$$ | . . X . . . . . . . . b . . . O O O . |
$$ | . . . X . . . . . . . . X . X . . . . |
$$ | . . . . . . . . W W W W X . . . O . . |
$$ | . . . . . . . . W X X X O O O O . X . |
$$ | . . . . . . . . . . . . X . X . . O . |
$$ | . . X . . . . . W . X . X . . . . O . |
$$ | . . . . . . . . . . . . . . X X X X . |
$$ | . . O . X . . . . . . . O O . O . . . |
$$ | . . . a . . . . . . . . X O . . . O . |
$$ | . . . . . . . . . . . X . X . O . . . |
$$ | . . . O . . . . . X . . X . . O . O . |
$$ | . . . . . . . . . . . . X . X X O . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I thought a move to strengthen the centre group was pretty much essential at this point. But no, Otake fearlessly played a! KataGo does indeed prefer a move in the centre (and has several suggestions, not a single clear favourite), but thinks that the game move is still playable (white's lead drops to 2 points). In many variations, white has b as a saving move.

Elf also prefers to play in the center rather than at a. In fact, it prefers b over a by 15½%.

xela wrote:That is a "+11% move" for my KataGo 1.3 setup. That is, KataGo doesn't find the move, but when I put it on the board, the winrate is 11% above KataGo's first choice. Also, is +8% (KataGo just looks at the clamp instead), and is +4%.

FWIW, is Elf's top choice, with 31.6k rollouts, and for it prefers the push through at D-11 with 61.5k rollouts, but only by 0.1%, i.e, not really at all. Since gets only 3 rollouts, it inherits its winrate estimate from , which is also Elf's top choice, with 21.7k rollouts. As mentioned before, was not Elf's top choice, but it gets a winrate estimate 13½% better for White than the top choice. OC, with only 1 rollout, it inherits its winrate estimate from , also Elf's top choice with 14.2k rollouts.

xela wrote:With these apparent blind spots, I agree that we can't trust KataGo's evaluations of other moves round about this part of the game. In particular, I don't think we can conclude that is actually a blunder.

Well, Elf thinks that loses 36% to par. Even if we think that its evaluations hereabouts are off by 13% in favor of Black, and if we assume that all the winrate difference between and its top choice is explained by that amount and so subtract 13% from 36%, we still get a loss of 23%, which is still a blunder in my book. (We cannot conclude that and and their evaluations are blindspots for Elf, given its number of rollouts.)

xela wrote:In hindsight, after seeing I wondered if was a mistake. But any alternatives appear to be even worse.

Is there a way to promote in Elf's searches?

I noticed that BQM265 is from an Otake game of the same period.

The book Relentless talks about some positions being "wide", in the sense of having lots of options. The BQM 265 position is very wide in this sense. In many variations, KataGo explores a lot of moves and is slow to settle on a clear-cut best play.

Click Here To Show Diagram Code

[go]$$Bc Suiuchi-Otake 1974-05-01, black to play
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . X X X . X . . . . O . . O . . X . . |
$$ | . X O O O . . . . , . . . . . X . . . |
$$ | . O O . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . d . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . a , . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . b . . . c . . |
$$ | . . . . O . . . . O . X . X . . . . . |
$$ | . X X . . . . . . . . . . . . X . . . |
$$ | . . . O . . . . . . . . . O . . . . . |
$$ | . . X O . . . . . O . . . . . X X . . |
$$ | . . X O . . . . . . . . . O . X O X . |
$$ | . . . . . . . . . . . . . . . O O O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

There are four "hot" areas of the board. On the left, can black invade at a, or can white get sente to defend there first? On the right, b, c, d and a few other places could be promising invasion points. At the top, normal play would look like this:

Click Here To Show Diagram Code

[go]$$Bc
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . b O . . . |
$$ | . X X X . X . . . . O 2 6 O a . X . . |
$$ | . X O O O . . . . , . 1 4 3 . X . . . |
$$ | . O O . . . . . . . . . 5 . . . . . . |
$$ | . . X . . . . . . . . . . . . X . . . |[/go]

But with the rest of the board looking so interesting, both players are keen to get sente, so white might tenuki after , or even after , and decide that allowing black a followed by b is a fair trade for something on the left or right side.

At the bottom, a hane by either player is big but gote. How does this compare with the rest of the board?

Click Here To Show Diagram Code

[go]$$Bc Black has first move at the bottom
$$ | . . . . O . . . . O . X . X . . . . . |
$$ | . X X . . . . . . . . . . . . X . . . |
$$ | . . . O . . . . . . . . . O . . . . . |
$$ | . . X O . . . . . O . . . . . X X . . |
$$ | . . X O . . . . . . . . . O . X O X . |
$$ | . . 3 1 2 a . . . . . . . . . O O O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

If black gets there first, there is some aji at a

Click Here To Show Diagram Code

[go]$$Wc White plays first at the bottom
$$ | . . 4 , . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . 5 O . . . . . . . . . b . . c . . . |
$$ | . . . . O . . . . O . X . X . . . . . |
$$ | . X X . . . . . . . . . . . . X . . . |
$$ | . . . O . . . . . . . . . O . . . . . |
$$ | . . X O . . . . . O . . . . . X X . . |
$$ | . . X O . . . . . . . . . O . X O X . |
$$ | . 2 1 3 . . . . . . . . . . . O O O . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

After , it's harder for black to invade at because threatens to kill the corner. But KataGo says that black shouldn't be afraid, and shows sequences where black leaves a bunch of dead stones behind (either in the corner or at ) and still manages to achieve whole-board balance.

Some variations for your entertainment:



And here's the full game, showing how people used to play sanrensei before the AI revolution!