It is currently Sat Sep 21, 2019 10:47 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 7 posts ] 
Author Message
Offline
 Post subject: AI verdict on Dosaku
Post #1 Posted: Sat Aug 10, 2019 10:51 am 
Oza

Posts: 2352
Liked others: 15
Was liked: 3417
We now come to the final part of Ohashi Hirofumi’s mini-series on famous historical games re-assessed by using the Chinese Golaxy bot at no-komi settings. We humans are in for a real treat. I hope I will be excused an unusually long post. This is not just one present under the Christmas tree. You get to scoop all the presents there, but in addition every bauble on its branches contains a delightful sweet.

I have added lots of comments of my own, and references to Lizzie are mine, but it is Ohashi who is Santa Claus.

The game is 1684 - note the date carefully. It Honinbo Dosaku playing Yasui Shunchi at the highest level of a Castle Game. It became “Dosaku’s Masterpiece.” It was his own choice of the most satisfying game he ever played as White, even though he lost. It is a 2-stone handicap game, but in Shunchi he found an opponent he thought was truly worthy. He said of Shunchi that, “He is a first-rate player and need feel no shame before the ancients. He is a rarity even by past standards. In this game there was no manoeuvre or moves of his that were not worthy of some praise.”

From this game derived the practice of calling a 1-point loss a player’s masterpiece.

And Golaxy seems to agree with Dosaku’s assessment! Which means also agreeing with over three hundred years of pro opinion about just how good Dosaku really was. This gushing praise is where the appellation 13-dan for a super-duper player who can eat three Shredded Wheats for breakfast came from.

The figures speak for themselves, of course. From 1674, when he was 29, till his death in 1702 he never had to take White -and he rarely took Black even before then. On top of that, he scored 14-2 in Castle Games and the two losses were by just 1 point when giving two stones (one was the present game). On the face of it this might not seem quite as impressive as Shusaku’s longer winning streak in Castle Games, but Dosaku’s record is untainted by the allegations of game fixing that Yasui made against Shusaku (the Honinbo-Yasui rivalry still rumbling 200 years on!).

The mention of Shusaku turns on an interesting sidelight. About both Dosaku and Shusaku is told the story that they first heard the sound of go stones when in their mothers’ wombs. Recent theory on the brain and music in fact seems to suggest that children are attracted to sounds they hear in the womb. Maybe this includes go stones. But leaving that aside, it does seem likely that a child whose mother is an accomplished player (as in these two cases) is likely to start being taught go at a far earlier age, and perhaps with more dedication, than a child who learns from his father or child companions. This seems to tie in also with the common observation that the youngest child in a family of go-playing siblings is the one most likely to be outstanding. They simply end up learning earlier. Go Seigen is maybe the most famous example of this, but the phenomenon has been widely reported in both go and shogi.

Incidentally, as awesome as Dosaku is going appear, it should never be forgotten that he himself seemed a little in thrall to his pupil Ogawa Doteki, who was considered to be close to Meijin standard when he died at 21. Go Seigen also spoke very highly of this example of “those whom the gods love die young.”

More concretely as to why Dosaku was so strong, the usual opinion is that he is the non pareil example of a player who strives for maximum efficiency with each move. Since this is how almost everyone is characterising the style of the AI bots, the present comparison offers to be most fascinating. In fact, it is no spoiler to say that Dosaku passes the test with flying colours. A couple of points emerge that are even more intriguing.

At last, down to the game. It is a handicap game, but Golaxy was set up to cope with that, and in the starting position it rated the game as a Black win by 21~22 points, with a win rate of about 94%. Presumably that includes taking account of White playing first.

For comparison, Lizzie (= LeelaZero) starts with a win rate of about 25%. In general it seems to come up with the same sort of moves of Golaxy, but the win rate figures seem to go skewgee, slightly at the beginning but then wildly about half way through. It is, of course, trained on 7.5 komi. More on that later, but I have mostly avoided mentioning Lizzie for that reason in this Part 3.



We apparently see a discrepancy at once. Lizzie has Dosaku falling further behind for the early moves. Golaxy says he maintains the initial evaluation.

One difference is that in its candidate moves Golaxy preferred the press at F16 for Black 10. That’s a typical AI kind of move. In one of the many insightful throw-away remarks that Ohashi makes in all his writings, he says that this is probably linked with the bots’ liking for two-space shimaris (and presumably other centre-oriented moves). That sounds blindingly obvious once it is said, but I don’t recall seeing anyone saying it before.

The first big change in score registered by Golaxy was for Black 30, which led to a 3-point change. Golaxy prefers to sagari at H18. What Golaxy predicts will happen then is a moyo-isation of the upper right with Black leaving a weak group in the upper left. But by running away with this group, Black will not end up being chased into his own moyo. So this seems like a perfect example of the one-weak-group strategy.

Ohashi notes that just 32 rather than H18 was also fine, according to Golaxy, but the omission of Black 30 is not really an AI insight. Many human pros condemn this sort of move: its main drawback is that it strengthens White. However, Ohashi implies that Shunchi, though mistaken, had reasons for 30, and it was not simply a case of “monkey see atari, monkey play atari.” In fact, choosing to sacrifice 30 in the follow-up with 36 and 38 was “good judgement.”

White 39 was a surprising move that showed Dosaku’s flexibility, according to Ohashi. The obvious move for White is to push up at H13. But after one more forcing push Black will then nail down the corner with P17. The reason 39 was surprising is that Golaxy did not include it in its list of candidates (nor did Lizzie), but once it was played the score did not change. This phenomenon is so common with bots that I think we really ought to have a term for it. Any suggestions?

But another aspect that Ohashi highlighted is that, in the ensuing phase of the game, Golaxy showed no dubious moves for either side.

White 63, setting up a big ko in the comer, was the next crunch point. It ended with a big trade: 68 for 69. Ohashi, who said studying Dosaku’s collected games was the first major study he ever made on the way to becoming a pro, had long thought this trade favoured White. but Golaxy saw no change in the score for some thirty moves. In other words, the trade was completely even.

So, if nothing else, Dosaku was a better judge of the game than Ohashi!

Black 70 was premature, said Ohashi. Interestingly, the Golaxy score does not go down immediately after this, but Ohashi claims it is the reason White can whittle down Black’s lead later.

But most interesting is that Go Laxy preferred N10. and his cousin Go Seigen preferred N11, though in both cases only after inserting the probing exchange O4 for O3. And Seigen said it first! Several of us have noted on this forum that we have seen similarities with AlphaGo’s plays in Go Seigen’s games. This seems like concrete proof, and there are other examples in the Golaxy analysis.

Just for the record, Hayashi Genbi and others have, in the past, defend Black 70. They may well have been wrong, but at least it has been recognised by humans that this was a contentious point in the game. A bonus point for us!

Black 76 was a clear mistake, apparently, and another case where Go Seigen agreed with his silicon counterpart. The significance of that, according to Ohashi, is that Shunchi’s move looks “obviously” good. It makes miai of capturing some stones to left or right. But the numinous experts Golaxy, Go Seigen and (by implication from the follow-up) Dosaku all say Black should force at 77 before playing 76 (i.e. losing the miai capture).

Dosaku’s follow-up was to sacrifice the corner with 77: good judgement, because the sabaki initiated at 79 is excellent compensation. In fact , Golaxy says it has reduced Black’s lead by about 3 points, and it has now been cut to 14~15 points.

The sequence up to 86 is essentially forced, but Golaxy made what Ohashi considered a surprising suggestion of the force at L14 then B13. That is, 86 was not bad but L14 was the top candidate.

White 87 was the highlight of the game. Shining enough for Ohashi to say it twice! He would have played 94. But Dosaku’s move was a prime example of striving for maximum efficiency - and Golaxy completely agreed with Dosaku (there is, of course, some tactical reading also involved). Whether Go Seigen would have also played this move is unclear, but he does praise it as a “strong move typical of Dosaku,” and if we assume (as I think we can from Go’s explanation of his opinion) that a strong move means one striving for maximum efficiency, we can say all three are on the same page.

I’m going to go off on a short tangent here. We saw earlier that 77 was good because it allowed White to turn to sabaki at the top. The reason 87 was so good is that it continues this sabaki manoeuvre efficiently. If you are of the school that thinks sabaki means light play, this may surprise you. In fact sabaki just means ‘coping’ and can be heavy or light. I think that matters because words translate, via associations, into concepts in our minds and if we just have the concept of light sabaki in there, we are not likely to make the heavy sort of move that Dosaku and Golaxy came up with.

But there’s a little more to it, I believe. One common association with ‘heavy’ is overconcentration – something to be avoided. But overconcentration can be accepted if it overconcentrates the opponent even more. One thing that struck me about Ohashi’s write-up here was his much repeated use of the phrase “whittling down [Black’s] lead” (sa o tsumeru). It’s not an unusual phrase in ordinary Japanese, but it’s on the rare side in go. I may be overreading it, but I did wonder whether Ohashi was trying to make the point that Dosaku’s style of seeking maximum efficiency implied seeking to whittle down the opponent rather than going for the big kill (or the big territory).

Furthermore, when Ohashi studied Dosaku’s games early in his career, what he latched on to most was his new ideas, and sabaki above all. Modern players tend to dismiss the old masters because they didn’t know the latest joseki or fuseki, and even now it is the new joseki lines that most people are gawping at when looking at AI games. Perhaps they should be looking instead at the similarities in open play between masters like Dosaku and Golaxy.

Which brings us back nicely the game. Golaxy agreed with Dosaku that 97 was best.

Black 98 was another slip by Shunchi, but Go Seigen again had pointed this out before Golaxy came along.

102 was also a small loss, and for the first time Black’s lead slipped below 10 points.

White 107 was the final element in the success of Dosaku’s sabaki manoeuvre. Ohashi reminds us again that Dosaku has made good use of his whittling knife despite the fact that Shunchi has made no egregious errors.

But he was an a downward path, and his sequence of 108~114 saw his lead dwindle to 5 points.

Black 120 starts a fight. Golaxy agreed that that was one of the candidates to consider, but it seems risky and the bot recommended B13 instead. Similarly, the bot prefers 128 to be the cut at 129. The life & death and other calculations around here are very tricky and may be just too hard for humans.

White 133 starts another ko. Although I have just made a reference to dauntingly complex calculations, it seems worth noting that Dosaku’s so far impeccable play has included complex ko fights. At this point Black was just 3 points ahead. More “whittling,” it was noted.

But even Homer nodded, and with White 137 Dosaku showed at last he was human. And good for him, I say!

He did not play a bad move. He just held back. For once, he did not strive for maximum efficiency. For 137 he should have played one line further at 138, for example.

It would be possible to imagine this change as being due to extraneous factors such as tiredness. But the practice of shitauchi – the practice of playing the game at home a couple of days before the Castle Game ceremony so as just to show the bored Shogun the highlights – had begun just before in 1672, so we can’t be sure the fatigue of playing a whole game in a day was a problem. If you are a conspiracy theorist, or have just read too much [i]Hikaru no Go[/], you might also wonder whether Dosaku had embarked on a plan to ensure a 1-point game. But Ohashi comes up with a much more interesting possible explanation.

His point was that the fault with 137 ultimately lay in Black getting the “thick” move of 144. This interested me instantly, because one thing I have noticed strongly in all the many commentaries I have been through lately (almost 100 games in the Genjo-Chitoku book, many more than that in Shuei’s games, and the games of the forthcoming Go-Iwamoto book, and all based on multiple commentaries – in other words, LOTS of data) is the very high usage of “thick” in praising or recommending moves in the oyose – the large boundary-play stage, or macro-endgame. What I have noticed above all is that the term is used a lot but an explanation is never given beyond “thickness.”

Well, despite us generally saying that the problem with bots is they can’t speak or explain like humans, Golaxy is the first to give me some specifics of what is meant: Black’s “thick” move 144 made his lead go back up from 3 to 7 points. That’s a lot – more than I expected – and showed me why the pros kept stressing thick endgame moves. (Oyose is normally reckoned to start when plays are at the 12-point value, though some pros set the bar higher.)

However (remember you get lots more presents than expected from this Part 3!), Ohashi put the icing on the cake by revealing that he had come to realise, mainly from his study of AI go, which of course goes back before AlphaGo days, that the oyose is probably the most difficult phase of the game, and the one where pros can perhaps learn most from bots. O Meien has made similar remarks about the difficulty of the oyose.

And in that vein it will be no surprise that human Dosaku made his first real mistake in the oyose, with White 155.

After 169, Black’s lead was 4 points, but Black still had opportunities to shoot himself in the foot. And he did. 180 was a mistake (and 185 was, too). With an alternative play he could have apparently taken his lead back up to 10 points, which seems to imply that Dosaku must have boobed somewhere but it is not clear where – was this the fault of 155? In general we don’t get to see data for all that many moves, nor we do we see how close the human is in those cases where the bot prefers another move, The overall impression, though, is that Dosaku has done astoundingly well and Shunchi has also covered himself in glory.

It seems as if Dosaku made just one really serious mistake, and einmal ist keinmal, as the Germans wisely say.

At any rate, Ohashi says the analysis confirms that the game can be considered a masterpiece. You can make an assessment of sorts for yourself by looking at the graph below [see Edit note below - technical issues with forum software; graphs are shown in wrong order]. It shows the changes in evaluation by Golaxy, with Black’s lead in points on the y axis. Obviously the x axis shows the number of moves.

Attachment:
IMG_0057.PNG
IMG_0057.PNG [ 110.86 KiB | Viewed 889 times ]


The Lizzie version of this graph is very different. At about the halfway stage it starts to go up and ends in more or less a large V shape.

For comparison, here below is another Golaxy graph, from the Shusaku-Genan game (no graph was given for Jowa-Intetsu). In this case, however, the y axis shows the % win rate for Black.

Attachment:
IMG_0058.PNG
IMG_0058.PNG [ 100.25 KiB | Viewed 889 times ]


Jowa likewise seemed to score well in comparison with the computer in this mini-series, but this Part 3 in particular seems to raise some perplexing questions.
If Dosaku, despite living over 300 years ago, could perform so well on the Golaxy chart, can we really say modern players are stronger than him? Was he really 13-dan? And was Go Seigen? I have a memory of reading some research recently that, I think, showed, on the basis of AI research in chess, that Capablanca may have been the best player ever – certainly up with modern counterparts. Have people been dazzled too much by joseki or opening innovations in go and chess instead of focusing on the deeper essentials of the game, such as sabaki, overconcentration, and efficiency in general?

If Dosaku is so good and plays like a bot, is he nt the best model for modern players to study? We have enough games – almost 160.
And although even humans like Dosaku, Jowa and Go Seigen may nevertheless not be as good as the bots over an entire game, are they close enough to copy and trust? That very question seems, to me, to suggest a way to study. If we pit the best humans against the bots but insist that N number of times the bot must make a random move (or choose the lowest move on its candidates list, or whatever), to give the humans a winning chance in an even-game environment, will we be able to see at which stage the humans mainly go wrong (e.g. the oyose, as Ohashi speculates), or how many errors can be tolerated – or what kind of errors?

In any event, there is a lot of stimulating thought to be gained from Ohashi’s series, and I would like you consider even more from him, either in his many previous articles in Go World or, even better, his latest books.

An enjoyable one is Igo AI Jidai no Shin Fuseki-ho from 2017 (ISBN 978-4-8399-6381-1. Here, Ohashi does not just present the New Fuseki aspects of modern bot games but attempts a retrospective of and synthesis with the 1930s New Fuseki (with lots of stuff on tengen). In Japanese, but heavily diagrammed and with not too much text.

From 2018 there is a unique book called AlphaGo Teach Kanzen Gaido (ISBN 978-4-8399-6628-7). This complete guide to using the AlphaGo Teach tool available on the internet. It covers joseki and fuseki. I haven’t tried that tool so my comments on the book are worth little, but Ohashi is a writer you can trust. Again in Japanese but heavily diagrammed. If you already know how to use Teach, ypu may be able to breeze through this book. The examples look rather different from those I’ve seen on this forum, and they cover a wide range of bots

Another book for the more seriously minded is Igo AI Shin Jidai by O Meien (ISBN 978-4-8399-6254-8) a survey for layman of all aspects ote recent AI revolution in go. It went through its first two editions in 2017. It is in Japanese but heavily text oriented this time, and I also find O’s writing style is like wading through molasses. But you can’t deny the quality of his thinking and special insights.

If you’ve made it this far, you probably feel like you’ve been throuh molasses , too. But Christmas is meant to be sweet and sugary, and I hope you’ve enjoyed it. Christmas in August? Well, why not? I bet that’s put the wind up the turkeys.

Edit note: I have tried multiple times to get the graphs in the right order. The program ignores me I have given up. Just be aware that the first graph should show a continuous downward slope and the second a kind of V shape.


Last edited by John Fairbairn on Sun Aug 11, 2019 1:04 pm, edited 7 times in total.

This post by John Fairbairn was liked by 13 people: Bill Spight, Bki, gowan, Hades12, HKA, jeromie, mycophobia, PeterHB, SoDesuNe, sorin, Umsturz, wolfking, zermelo
Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Dosaku
Post #2 Posted: Sat Aug 10, 2019 11:58 am 
Honinbo

Posts: 8820
Liked others: 2602
Was liked: 3009
John Fairbairn wrote:
Oyose is normally reckoned to start when plays are at the 12-point value, though some pros set the bar higher.


I assume you are using swing values. Based on the books I have read, I have always thought that yose started at higher move values. But one thing I have noted is that corner ko fights or plays where seki is one outcome often take place when swing values are around 14 or 15. So considering the yose stage to start when swing value are around 12 would eliminate most of those fights from the yose stage. (FWIW, my preference is to consider them to be part of the endgame. :))

----

BTW, your graphs seem to be in reverse order.

_________________
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 Dosaku
Post #3 Posted: Sat Aug 10, 2019 6:48 pm 
Lives in gote

Posts: 338
Liked others: 298
Was liked: 170
Thank you so much for these three "AI verdict" posts, John!!

John Fairbairn wrote:
One difference is that in its candidate moves Golaxy preferred the press at D16 for White 10. That’s a typical AI kind of move.


Is it really D16, or rather D15 (press)?

_________________
Sorin - 361points.com

Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Dosaku
Post #4 Posted: Sun Aug 11, 2019 5:29 am 
Gosei
User avatar

Posts: 1556
Liked others: 365
Was liked: 289
Rank: 2d-2k
Thank you very much for your three posts! They were certainly a treat worthy of christmas : )

_________________
My "guide" to become stronger in Go

Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Dosaku
Post #5 Posted: Sun Aug 11, 2019 7:32 am 
Lives with ko
User avatar

Posts: 181
Liked others: 4
Was liked: 56
Rank: EGF 5 dan
GD Posts: 29
Thank you for the excellent article. I took some time to set up the go board, print the text and also find the Appreciating Famous Games for a good Sunday afternoon go study session.

Now, even more inspired after going through the article, I'll check out my Dosaku collection.

Cheers,
Vesa

Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Dosaku
Post #6 Posted: Sun Aug 11, 2019 12:34 pm 
Dies in gote

Posts: 23
Liked others: 16
Was liked: 2
OGS: DeprivedWaste
Very interesting! Thanks for taking the time to post this here, I'll def. check out Dosaku's game in the future, if even contemporary AI seem to agree a lot with his moves

Top
 Profile  
 
Offline
 Post subject: Re: AI verdict on Dosaku
Post #7 Posted: Mon Aug 12, 2019 9:46 am 
Lives with ko

Posts: 156
Liked others: 9
Was liked: 11
Rank: 2D
Tygem: shiva
I feel as if I have disrespected the old Japanese greats by not studying their games. Dosaku amazes me. The strongest aspect of his play is what I would consider "timing" and using aji. If you play the correct move, but your timing is off, you ruin a sequence. Move 39 is the first example of Dosaku using perfect timing. If he doesn't play it now, he may never get it, and he risks allowing black to become solidified in the upper right. To go with the idea of "timing", I think we can all agree that Dosaku is patient.

I agree with Ohashi's thinking of the big trade of 68-69 favoring white. If what I understand about Golaxy is correct, the bot is keeping score. So while the point exchange may have evened out, with white taking the upper left and black taking the upper right, I think white gets the slightest advantage by having the O16 stones which limit blacks growth. It also sets up move 87 nicely. Getting the ko there was a sweet sequence for white, and I think his timing was perfect. Think about it. He gets the ko early in the game where black has very minimal threats. If black loses the ko, he loses the corner and he gets a weak group in the middle, and a huge cut at K13. I think black's play is forced because he has no threats. Another example of Dosaku using timing to his advantage.

In math, when you solve a problem, you have to give "proofs" that your solution is correct. I think white G4, at move 115, is whites proof that B6 was premature. I think this is yet another one of Dosaku's examples of timing and patience and using all of the aji on the board. The sequence up to white 119 gives white amazing shape on the outside and maximum efficiency of his stones which John refers to.

Amazing. I will be reviewing Dosaku's games after this post. :clap:

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 7 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users 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