It is currently Sat Sep 21, 2019 6:25 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 5 posts ] 
Author Message
Offline
 Post subject: AI verdict on Shusaku
Post #1 Posted: Thu Aug 08, 2019 6:04 am 
Oza

Posts: 2352
Liked others: 15
Was liked: 3417
Having digested Ohashi Hirofumi's review of Golaxy's full-day AI assessment of Jowa's Three Brilliancies game (see separate thread), I turned to Part 2, Shusaku's famous game with Genan Inseki. It is famous because Shusaku's 127 was dubbed the Ear-Reddening Move when a doctor saw Genan's ears redden upon seeing it.

Spoiler alert 1: if you are a Shusaku fan you will be disappointed. Spoiler alert 2: if you trust your doctor, maybe think again!

In broad terms, what we see here is similar in many respects to the Jowa-Intetsu game. The nominally stronger player does in fact make far fewer mistakes. Golaxy and Lizzie strongly tend to share opinions, though somewhat more differences emerge here. And psychology seems to play a big part in human pro play.

A reminder that Golaxy is special in this exercise because it can cope with no komi and can give point evaluation (and assesses that on average Black wins by 6.1 points from an empty board), and that the references to Lizzie are my own.

At the time of this game Shusaku was just 17 and 4-dan. That would equate to a 1-dan pro nowadays, as the first three dans were then also given to amateurs. Genan was Vice-Meijin but perhaps no longer at the peak of his powers and was in semi-retirement in Osaka. Part of the reason this game became celebrated was precisely because it represented an East-West Japan duel - a feud that even still reverberates today.

What happens on the board here is certainly not enough to make the game famous. Shusaku was the main culprit as regards the number of bad moves. Genan made far fewer but perhaps worse ones, and (my observation, not Ohashi's) made the most egregious mistakes with his moves immediately after game resumptions. In those days, there was no sealed move. When it was time to break for the day, White decided when to stop, and obviously chose a point at which Back had made an "interesting" move. In thise case, the first break was on move 89, and play was not resumed for three days. So Genan had ample time to concoct a perfect reply. At that point the game was close to even, but after White 90, Black's lead shot up. Similar things happened after Black 141 and 189, the other two pause points - not just small mistakes but clangers, and clangers of an ilk that did not occur in normal play.

It is entirely my own speculation, but based on known patterns of behaviour in other games, what may have happened is that master Genan relied on his pupils to carry out adjournment analysis. This could have seemed like a big advantage as the game was in Osaka, where Genan had lots of pupils who could beaver away and tell him what to play. Shusaku was on the road and had to rely mainly on himself. In short, Genan may have been complacent, lazy or too trusting.

Apart from the adjournment clangers, few mistakes by Genan are mentioned. They are also relatively minor. For example, his choice of the taisha on move 10 rather than the press is queried. In the resulting joseki, incidentally, Shusaku's move 25 was labelled bad and should be the slide at 29 - but this is a point that has already been made by human commentators.

Black 37 was Shusaku's first major slip. The win rate for Black starts at around 70% in his favour, but after this move it fell to 60% according to Golaxy, a Black win by 1 to 2 points. (Lizzie, trained on 7.5 komi, puts the drop at just 4 percentage points).

Black 53 was also awful according to the bots, but has been regarded as a famously good move by Shusaku fans. Ohashi's comment is insightful here. Although Shusaku's fame rests to a large degree on his style of storing up strength early on (his favourite "Shusaku kosumi" is the best known example), Ohashi believes his true genius lay in his sensitivity to the ebb and flow of the game. This is a concept that has its own label in Japanese. The concept is, I'm sure, easily recognisable to western players but perhaps lacks a specific label. In Japanese it is 勝負勘 or just 勘 (kan). It reflects the fact that humans lack the ability to read things out very deeply and so must instead rely on a kind of intuition. Black 53 (rather than the usual block) was an example of this kan. But of course a vague feeling is not enough - you also need the skill to act on it with tricky moves. If you have read the Jowa thread, you will recall that Ohashi made a similar point about elite players such as Ke Jie and Iyama Yuta.

Genan's first real mistake not to do with adjournments was White 60. However, he got away with it because Shusaku made an equally bad reply. However, this Black reply was an interesting case of Black storing up strength à la Shusaku and in a way that Ohashi said that appealed to him, too - a move he could easily imitate. Maybe it is such cases that will provide good study material for those wishing to unearth reasons why AI bots are stronger.

Black 79 was another mistake, then came the bad Black 90 already mentioned, and soon after Black 95 was an awful missed chance (if he had played K18, his win rate would have been 78.5%). 97 was a loss of a couple of points, too.

At long last we come to the auris rubicundula, as we medical men say. On move 125 Black had a win rate according to Golaxy of 40%, that is White was going to win by 3-4 points. Genan then played 126, which the bots say he should have just omitted (i.e. he should go straight to 128) and kept his lead intact.

Black 127 was the move at which the Osakan doctor predicted Genan would lose as he deduced from a reddening of the ears that Genan had been fatally wounded. There were several problems with that prediction. For one, 127 was not a specially good move - not actually bad but not good enough even to register in the list of candidates drawn up by either Golaxy or Lizzie. For another, White was not yet losing. And for yet another, this game of wild fluctuations was about to fluctuate hugely both ways. If it hadn't been for the fact that Hiram Codd did not invent fizzy drinks until 1875, we might say the good doctor's opinion was codswallop. At any rate, two things seem more likely. One is that Genan's blush was one of annoyance with himself for wasting the 126-127 exchange. The other is that he'd neem on the wallop himself.

Genan's real problem was not 127 but another adjournment mistake (more wallop?). 142 was BAD. Shusaku's 143, in reply, has been praised by humans as good. Golaxy agreed. Lizzie, however, seemed totally unimpressed by it.

The next major change was at the next adjournment. Genan's 190 was VERY bad, turning a forecast win into a 3-point loss. Yet Black managed to lose even this late lead! At move 213 the forecast was for a jigo. In the end, Shusaku won by a slender margin, not having really covered himself in glory. He was a big hero to the Honinbos, of course, but that was for other reasons.

There was a problem with Golaxy's suggested play in the endgame. A style of play in that, when comfortably ahead, it eschews simple, slow moves in favour of trying to settle the game instantly and never mind the complications, which do bite it in the bum occasionally (hooray!). Having seen O Mein's masterful analysis of this endgame, Ohashi pointed this "sickness" - kimetagari-byo - out to the Golaxy team, who confirmed it as a known problem. As he said, even Golaxy is not yet a god. And Shusaku certainly is not, Hikaru no Go notwithstanding. Will Dosaku (Part 3) become our numen, as we sacerdotal men say?



[admin] John, I hope you will pardon me for editing your post. I found that it made much more sense if I had a copy of the game to look at. So I added the SGF below. -Joaz [/admin]



This post by John Fairbairn was liked by 13 people: Bill Spight, explo, gowan, HKA, jeromie, Joaz Banbeck, mycophobia, silviu22, SoDesuNe, sorin, Umsturz, wolfking, zermelo
Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Shusaku
Post #2 Posted: Thu Aug 08, 2019 7:36 am 
Lives with ko

Posts: 156
Liked others: 9
Was liked: 11
Rank: 2D
Tygem: shiva
Thank you for this write up. Golaxy is incredibly interesting, due to its emphasis on points. As we see from your write up, black at one point had a 70% win chance and then fell to 40% win chance. But as we can see, this huge swing is only due to a couple of points on the board. This is why in the amateurs we can see a 90/10 victory chance get flipped so easily, because we often see 20 point or more exchanges late into the middle game. I'm interested in using top pro games of the modern era and seeing the point swing, as well as cross referencing that with the top amateurs to try to establish a range of errors. I have always thought of getting stronger in go as not really 'getting stronger' but limiting your mistakes. Getting stronger at go is attempting to get closer to perfect level of play. An amateur shodan may make 3-4 large mistakes and dozens of small ones. Maybe a 5D makes 2 large mistakes and a half dozen small ones. Once we get to pro level, notwithstanding a reading mistake, we see mostly small errors with maybe 1 large mistake. Basically, what is the margin of error in go at a pro and amateur level? That's a huge research topic right now.

Thanks John

Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Shusaku
Post #3 Posted: Thu Aug 08, 2019 8:27 am 
Tengen

Posts: 4801
Liked others: 0
Was liked: 648
Hades12 wrote:
An amateur shodan may make 3-4 large mistakes and dozens of small ones. Maybe a 5D makes 2 large mistakes and a half dozen small ones.


Not for long thinking times. There the large mistakes are ca. 0.1 - 1 on average for 5d. Maybe your numbers apply to online fast games.

Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Shusaku
Post #4 Posted: Thu Aug 08, 2019 12:01 pm 
Honinbo

Posts: 8820
Liked others: 2602
Was liked: 3009
John Fairbairn wrote:
Having digested Ohashi Hirofumi's review of Golaxy's full-day AI assessment of Jowa's Three Brilliancies game (see separate thread), I turned to Part 2, Shusaku's famous game with Genan Inseki. It is famous because Shusaku's 127 was dubbed the Ear-Reddening Move when a doctor saw Genan's ears redden upon seeing it.

Spoiler alert 1: if you are a Shusaku fan you will be disappointed. Spoiler alert 2: if you trust your doctor, maybe think again!


Well, the doctor's evidence was earsay. :lol:

John Fairbairn wrote:
A reminder that Golaxy is special in this exercise because it can cope with no komi and can give point evaluation (and assesses that on average Black wins by 6.1 points from an empty board), and that the references to Lizzie are my own.

{snip}

What happens on the board here is certainly not enough to make the game famous. Shusaku was the main culprit as regards the number of bad moves. Genan made far fewer but perhaps worse ones, and (my observation, not Ohashi's) made the most egregious mistakes with his moves immediately after game resumptions. In those days, there was no sealed move. When it was time to break for the day, White decided when to stop, and obviously chose a point at which Back had made an "interesting" move. In thise case, the first break was on move 89, and play was not resumed for three days. So Genan had ample time to concoct a perfect reply. At that point the game was close to even, but after White 90, Black's lead shot up. Similar things happened after Black 141 and 189, the other two pause points - not just small mistakes but clangers, and clangers of an ilk that did not occur in normal play.


Elf, laboring under the double handicap of assuming a 7.5 komi and relatively fewer playouts than Golaxy, takes a different view of the play after :b89:. It thinks that :w90: should have been the hane at :w96:, as well as :w92: and :w94:. It rates :w90: as losing more than 10% in winrate, but :w92: and :w94: are even worse! However, it thinks that :b93: and especially :b95: are whoppers, as well. When the smoke clears after :b97:, Elf rates Black as worse off than before :b89:.

Elf thinks that White 142 is also a whopper. By this time, with red ears, perhaps Gennan was playing to induce a mistake by Shusaku. Elf also thinks that White 190 was very bad, but also that Black 191 returned the favor in spades, losing much more than White 190. Elf recommends the simple connection to the peep. :) If Gennan was playing for an error, he certainly succeeded with White 190. According to Elf.

John Fairbairn wrote:
Genan's first real mistake not to do with adjournments was White 60. However, he got away with it because Shusaku made an equally bad reply. However, this Black reply was an interesting case of Black storing up strength à la Shusaku and in a way that Ohashi said that appealed to him, too - a move he could easily imitate. Maybe it is such cases that will provide good study material for those wishing to unearth reasons why AI bots are stronger.


Elf thinks that :w60: was horrid, losing over 40% in winrate. It thinks that :b61: was also a blunder, but losing only 20% or so. It thinks that both plays would have been better at H-03.

John Fairbairn wrote:
Black 79 was another mistake,


Followed by an error by White, but not as bad, according to Elf. Neither player seemed to appreciate the importance of M-11.

John Fairbairn wrote:
At long last we come to the auris rubicundula, as we medical men say. On move 125 Black had a win rate according to Golaxy of 40%, that is White was going to win by 3-4 points.


Elf, still assuming 7.5 komi, estimates Black's winrate at 4%. If Golaxy's territorial estimate is correct, that would amount to around an 11 pt. loss at this point in the game.

John Fairbairn wrote:
Genan then played 126, which the bots say he should have just omitted (i.e. he should go straight to 128) and kept his lead intact.

Black 127 was the move at which the Osakan doctor predicted Genan would lose as he deduced from a reddening of the ears that Genan had been fatally wounded. There were several problems with that prediction. For one, 127 was not a specially good move - not actually bad but not good enough even to register in the list of candidates drawn up by either Golaxy or Lizzie. For another, White was not yet losing.


According to Elf, Black 127 returned his winrate to 4%, compensating for White 126. Black should have crawled to C-18.

John Fairbairn wrote:
Genan's real problem was not 127 but another adjournment mistake (more wallop?). 142 was BAD. Shusaku's 143, in reply, has been praised by humans as good. Golaxy agreed. Lizzie, however, seemed totally unimpressed by it.


Elf thinks Black 143 lost 3% to par, surely within the margin of error. It prefers the atari at E-17, and after White replies, play in the center.

John Fairbairn wrote:
The next major change was at the next adjournment. Genan's 190 was VERY bad, turning a forecast win into a 3-point loss.


By this late in the game, both Lizzie's and Elf's winrate estimates are not worth much, I suppose. But for comparison, Elf thinks that White 190 lost 23%, from a Black winrate of 34% to 57%. :shock: (The idea that Black should probably win, giving 7.5 komi, does not square with winning by only 3 pts. with no komi. ;))

Quote:
Yet Black managed to lose even this late lead!

Elf thinks that Black 191 lost 49½%, blowing the lead right away.

Quote:
At move 213 the forecast was for a jigo.

For comparison, Elf gives a Black winrate estimate of 1.2%.

John Fairbairn wrote:
There was a problem with Golaxy's suggested play in the endgame.

Only a surprise to me because of the huge number of playouts per move for analysis. What was the goof? Or goofs? Thanks. :)

Edit: Elf suggests a bizarre variation at 226 ( :w26: ). White fills the Black liberty at R-06. :b27: at M-12 instead of connecting at T-11. :o :w28: at N-12 instead of atari at T-14. :shock: I can see suggesting a move that invites Black to make a mistake, but then not to follow up? Maybe it's a problem of whole board reading, even with 5,000,000 playouts.

[admin] John, I hope you will pardon me for editing your post. I found that it made much more sense if I had a copy of the game to look at. So I added the SGF below. -Joaz [/admin]


_________________
The Adkins Principle:

At some point, doesn't thinking have to go on?

— Winona Adkins

I think it's a great idea to talk during sex, as long as it's about snooker.

— Steve Davis

Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Shusaku
Post #5 Posted: Tue Aug 13, 2019 8:22 am 
Beginner

Posts: 7
Liked others: 2
Was liked: 1
John Fairbairn wrote:
In the resulting joseki, incidentally, Shusaku's move 25 was labelled bad and should be the slide at 29 - but this is a point that has already been made by human commentators.


There is a commentary in INVINCIBLE that Shusaku faced it for the first time. This is wrong. He played the "bad" move 25 intentionally. He used it just three months ago (1846-05-29) against Nakagawa Junsetsu in Osaka. He might have had his reasons to use it against Genan...

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 5 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: Google [Bot] and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group