Life In 19x19

Isn't there a section of Kageyama's "Lessons in the Fundamentals of Go" (or was it another book) where he looks at the following position (the black other corners might not be 4-4s, but top right is that shape). IIRC he says black's tenukis preceding 7 are ok, but allowing white 8 is bad because white gets a thick result without bad aji, so black should play the hane at a first to stop white getting the perfect shape so you have some aji for later and you can then take sente for the empty corner. And allowing white 8 is so bad the resulting whole board position is good for white.

Click Here To Show Diagram Code

[go]$$B
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 6 . . . |
$$ | . . . . . . . . . . . . . . 8 1 4 . . |
$$ | . . . 3 . . . . . , . . . . . , a . . |
$$ | . . . . . . . . . . . . . . . 2 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 7 . . . . . , . . . . . 5 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I can certainly understand the idea of avoiding the perfect shape for white and that's a good lesson of making bad aji in sente, but the whole board judgement that the above is good for white is not something I've really been convinced by (at various times over the years I've preferred black and white but never with a strong conviction). Yes white is super strong, but black is speedy (a tewari argument is white made a shimari with 4 then 2, rather conservative, then added another stone at 8, very conservative/slow, and then black made a dumb wedge at 1 that strengthened white but that bad exchange can't really be worse than white's 2 slow moves can it?). Like the 3-4 high approach knight move hanging connection old joseki, I've got the feeling teachers/books have a "trust me young padawan, once you are stronger you too will appreciate the power of the force slow but thick shape". Similar with some of Shuko's famous turns (e.g. example at https://senseis.xmp.net/?Turn). But what do the bots think? Elf v1 agrees that playing the hane is better for black at 89%, but black is already very good and playing the empty corner for 7 is 85% and not a good whole board position for white. LZ #157 is far less critical: it thinks taking the empty corner and allowing thick atari gives black a small advantage (52%) compared to the empty board (46.5%) and agrees the hane is a little better (54%).

P.S. If you shift the entire top right corner down a line (so it started as a black 4-4 then 6-4 approach, attach etc, obviously black wouldn't tenuki that but hypothetically...) then you end up with a shape like an AI/O Meien big high shimari with a dumb attach inside that got laddered making the opponent stronger, and with that result of a bigger corner Elf prefers white slightly (52%).

Click Here To Show Diagram Code

[go]$$B Oh, yeah? Tewari.
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . 6 . 2 . . |
$$ | . . . 3 . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . 4 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 1 . . . . . , . . . . . 5 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Does anyone think that Black is not better by a good bit?

Click Here To Show Diagram Code

[go]$$B Oh, yeah? Tewari.
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 8 . . . |
$$ | . . . . . . . . . . . . . . O 7 O . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Yes, this exchange is bad for Black. But surely only by a few points. Won't the White winrate of John Fairbairn's LZ be around 50% - 51%?

Edit: I reacted to the diagram and whole board evaluation without reading Uberdude's whole text.

Click Here To Show Diagram Code

[go]$$B Original diagram
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 6 . . . |
$$ | . . . . . . . . . . . . . . 8 1 4 . . |
$$ | . . . 3 . . . . . , . . . . . , a . . |
$$ | . . . . . . . . . . . . . . . 2 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 7 . . . . . , . . . . . 5 . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Yes, would have been better at "a". But that just means that was bad, and probably that was questionable, and maybe was, too. I defer to Uberdude and The Bots.

Won't the White winrate of John Fairbairn's LZ be around 50% - 51%?

Yes - closer to 50% perhaps. But what was interesting also to me was that if you take away moves 7 and 8, White's winrate drops dramatically to 38%.

Since 7 and 8 essentially represent Black giving White an extra stone, this seems to mean that a single stone early in the opening can be measured at something 12 percentage points. Does this mean anything significant?

Bill said (talking about the overconcentration of the turtle-shell): "If I'm right, the next 7 pts. of territory are worth around 8 Leela Zero percentage pts." Does this marry up with the above?

John Fairbairn wrote:

Won't the White winrate of John Fairbairn's LZ be around 50% - 51%?

Yes - closer to 50% perhaps. But what was interesting also to me was that if you take away moves 7 and 8, White's winrate drops dramatically to 38%.

Since 7 and 8 essentially represent Black giving White an extra stone, this seems to mean that a single stone early in the opening can be measured at something 12 percentage points. Does this mean anything significant?

Bill said (talking about the overconcentration of the turtle-shell): "If I'm right, the next 7 pts. of territory are worth around 8 Leela Zero percentage pts." Does this marry up with the above?

If you are talking about and in the tewari diagram, and Leela rates the position after in that diagram at 38% for White, then my estimations are off. I did not rate White's three stone super shimari as that bad, nor the loss to Black of the - exchange.

Leela Zero's winrate estimates with your settings (and that of others I have seen) strikes me as close to rating one extra stone in the opening at around 16% difference, which is consistent with the statistics for pros and amateur dans in the 1970s. Elf's winrate estimates are more extreme, which reflects its estimation of its own ability, I suppose.

If you are talking about and in the book diagram, we are above my pay scale.

John Fairbairn wrote: It's interesting, too, that in uberdude's investigion above, AI seems to be confirming we are right to put Fujisawa Hideyuki on a pedestal as another all-time great. Perhaps our judgements about the relative strengths of players are sound.

I wouldn't make any such conclusion yet, for starters I only looked at a few problems and haven't compared with problems of other pros. I explored a few more this evening:

- problem 65. Elf barely considers Shuko's solid connect suggestion, but it is at least better than the game move by a few percent. But both >10% worse than Elf's suggested tenuki to more open area of board. One of his aims is a subsequent forcing move and Elf agrees that's good and with some nice shape dodge moves in an example sequence.
- problem 66. Elf agrees with Shuko the game move was -12% soft (I got it right too, standard shape) and agrees with his moves for a longish sente sequence resulting in sacrifice of territory for thickness. But it thinks his gote defence at the end is a more significant (20%) mistake. Plus Elf says opponent wouldn't embark on the bad trade sequence but just dodge and play the local modern joseki.
- problem 67 (how to reduce a moyo). Elf says the game move Takao (white) played was not so bad and adjacent to the best move. Shuko's suggestion (in different area) was inbetween but his assumed sequence was indulgent reading with generally good white moves but many slack replies from black. Comment is "if 1 is exchanged for 2 it becomes a forcing move, doesn't it?" so he sees it as a good exchange but doesn't put enough effort into finding opponent's good resistance.
- problem 68 (continuation of moyo game with sabaki). Yuki played a soft (-12%) defence in response to a 2-space jump. Shuko wants to attach and cut it. Elf agrees cut is best, though gives extend as answer instead of hane to dodge the bad fight.
- 71 (how to answer a kick). In game the young pro headbutt adjacent stone (-10%) which Shuko criticised, he wants to play a fancy attachment inducing the descent move but Elf says it's terrible (-27%) and they aren't looking at the big picture and tenuki is best.

John Fairbairn wrote:
But the proverb that the turtle-shell capture is worth 60 points seems way off. To give just an example starting with 4-4, as below, White's winrate shot up to 64.2%.



I initially tried the centre point as the extra White move to balance the number of stones played. I assumed* that that was the best way to negate Black's influence. But LZ preferred A or B (both around 62%).

I thought the turtle-back thing was relevant assuming that the turtle back was created by capturing two of the opponent's stones - similarly, by capturing a single stone for the ponnuki. To me, that's a lot different than just playing a ponnuki or turtle back shape without capturing stones.

Just playing, e.g., ponnuki without capturing isn't that efficient - but capturing the opponent's stone is.

zermelo wrote: I’ll reiterate my earlier conclusion: If a much stronger player says that something is good, you can trust that it is good enough for your games.

Why?

Kirby wrote: I thought the turtle-back thing was relevant assuming that the turtle back was created by capturing two of the opponent's stones - similarly, by capturing a single stone for the ponnuki. To me, that's a lot different than just playing a ponnuki or turtle back shape without capturing stones.
Just playing, e.g., ponnuki without capturing isn't that efficient - but capturing the opponent's stone is.

If you count the stones in John's turtle example you will see 6 black and 4 white so presumably he did capture 2 white stones

Oh, sorry - my mistake. The position is still weird, because I prefer black here.

I guess there are quite a few white stones, but the turtle back gives me a lot of confidence.

As a side note, I also don't understand the seemingly implied conclusion that, because a bot is good at winning games, it implies that a bot is good at choosing moves from arbitrary positions. There are a couple of reasons I find this leap questionable:

1.) If a bot is "good enough" at evaluating board positions, a computer's search power can cause it to win games, even if it makes mistakes now and then.

2.) For any player, not or human, I'd guess that the quality of selected moves follow some distribution of quality, which isn't guaranteed to be uniform. For example, some moves may be super good and others may be just average or maybe even bad.

Given the above, I find this exercise useful for considering new alternatives, but it'd be a mistake to accept what the bots choose as gospel.

Maybe this is the same view that others already hold, but I wanted to express it clearly here.

That being said, a bot is more likely to be correct than me at any arbitrary position - I'll concede that

Kirby wrote:Given the above, I find this exercise useful for considering new alternatives, but it'd be a mistake to accept what the bots choose as gospel.

In general, accepting the move choices of bots as superior without understanding why they are better is a mistake. It's already hard to follow certain plays made by pros; I suspect that those who choose to blindly imitate them have seen limited success in the past. The reason why bots' moves work is because they look much deeper than humans and can hold many more better evaluations of intermediate board positions for further analysis; plus, said evaluations are based on pure experience, as opposed to human intuition, gathered from millions of games. And besides, we have already seen examples where certain moves made by pros were overlooked for several thousands of playouts by bots, yet once played, they seemed just as good if not better than the bots' alternatives. I've been meaning to rant on this for a while now and it just keeps brewing.

ponnuki without capturing isn't that efficient

Ponnuki without capturing is impossible. It's what the word means.

Given the above, I find this exercise useful for considering new alternatives, but it'd be a mistake to accept what the bots choose as gospel.

Absolutely - bots lose to other bots after all, and occasionally even to humans.

Kirby wrote:As a side note, I also don't understand the seemingly implied conclusion that, because a bot is good at winning games, it implies that a bot is good at choosing moves from arbitrary positions. There are a couple of reasons I find this leap questionable:

1.) If a bot is "good enough" at evaluating board positions, a computer's search power can cause it to win games, even if it makes mistakes now and then.

2.) For any player, not or human, I'd guess that the quality of selected moves follow some distribution of quality, which isn't guaranteed to be uniform. For example, some moves may be super good and others may be just average or maybe even bad.

Also, 3.) a Zero bot isn't guaranteed to have seen positions very similar to this arbitrary one in the self-play corpus that it has learned from, especially if the position being evaluated is pretty wacky.

dfan wrote:

Kirby wrote:As a side note, I also don't understand the seemingly implied conclusion that, because a bot is good at winning games, it implies that a bot is good at choosing moves from arbitrary positions. There are a couple of reasons I find this leap questionable:

1.) If a bot is "good enough" at evaluating board positions, a computer's search power can cause it to win games, even if it makes mistakes now and then.

2.) For any player, not or human, I'd guess that the quality of selected moves follow some distribution of quality, which isn't guaranteed to be uniform. For example, some moves may be super good and others may be just average or maybe even bad.

Also, 3.) a Zero bot isn't guaranteed to have seen positions very similar to this arbitrary one in the self-play corpus that it has learned from, especially if the position being evaluated is pretty wacky.

The position doesn't even have to be all that wacky. IIUC, the more any bot is trained by millions of games of self play, the more it has been trained on high level positions played in its style. How many of those positions are going to be similar to those that I face at move 75? The stronger the bot gets, the less similar its training positions will be to my games, n'est-çe pas?

Currently, bots are being trained to play better. I.e., to "solve" go starting from an empty board with a 7.5 komi by area scoring. Some are being trained for handicap games and different komi, as well. Producing better bots for those conditions is a fine thing in itself.

However, if I am going to use a bot to help me make better choices from the positions that I face, I would like it to be trained on those positions, or very similar positions. At some point, for human positions at any given level of play, as bots trained on self play continue to improve, we will reach the point of diminishing returns for evaluating those positions, since the bots' self play will not produce similar positions. It may be true that over time, bots will approach perfect play from an empty board with 7.5 komi, but that play may produce few positions like those produced in human games. The main reason that they do so today is that humans are imitating the bots.

To train bots for the purpose of coaching humans perhaps it would be good to train them on games (self play is OK, I think) starting from various positions taken from human games. If each training regime starts with sampling human games from go servers, then training positions will not diverge over time from those faced by humans.

The position doesn't even have to be all that wacky. IIUC, the more any bot is trained by millions of games of self play, the more it has been trained on high level positions played in its style. How many of those positions are going to be similar to those that I face at move 75? The stronger the bot gets, the less similar its training positions will be to my games, n'est-çe pas?

But these deep neural networks are good at this kind of generalization. Even if you create an unnatural position (but equal and not a 120 moves deep tsumego), it will still play sensible moves.