Learning from the Master

Uberdude · Post by **Uberdude** » Mon Jan 09, 2017 7:18 am

Another interesting trait of Master is that if you approach its 4-4, it answers high, 3-3 then after pushing down it likes to tenuki instead of choosing blocking on one side or another. This 3-3 invasion is very common in the Chinese opening these days and normally black splits from the approach (e.g. http://ps.waltheri.net/database/game/72524/), but in the AlphaGo self-play games released several months ago we saw it liked to block the side so a few pros tried that (and I saw some using it in the ama Gold Cup), e.g. http://ps.waltheri.net/database/game/73339/. Another nice idea is to play some approach moves elsewhere as probes and depending on how the opponent answers you choose which side to block (e.g. here white answers with low move on top side, so that becomes less interesting, so black blocks the right side: http://ps.waltheri.net/database/game/74171/). Ke Jie tried this in a game but his opponent just ignored the approach to directly defend with the tiger mouth shape (or hanging connection, not sure what to call it). He lost that game, and I wondered if that defence was too good for white: viewtopic.php?f=13&t=13513.

>> Edit, see post below, including the sgf here caused some rendering bugs.

Anyway, the recent Master vs Ke Jie game above also saw this shape and Ke made the corner tiger mouth, allowing black to make his own tiger mouth from the shoulder hit at bottom left. Now I can't help feeling if the corner move is slow. Quite a lot of humans seem to play it vs Master (I will include more later) and for sure it is a huge move, kind-of reverse sente, but it gives Master a chance to play a speedy and flexible opening. If the thinking behind it is "normally black blocks here so I will take advantage/punish Master by connecting here" then maybe it's wrong. In Ke's game he did later invade the top side taking advantage of the lack of a base of a 4-4, but Master managed the situation well and ended up trading and killing the invader.

Just looking at the local shape, I think diagram 1 is worse for white than diagram 2. The hanging connection does have some slight advantage with regards to the top side, but the thinner connection to the approach has all sorts of bad aji (things like a-d). There was a spectacular example of this in fact from another AI recently: viewtopic.php?f=10&t=13921.

Click Here To Show Diagram Code: [go]$$B diagram 1 $$ ----------------+ $$ . . . . . . . . | $$ . . . . . . O . | $$ . . . . . O . . | $$ . . X . X X O . | $$ . . . . . . d . | $$ . . . . . O . . | $$ . . . . . b . . | $$ . . . . . a c . |[/go]

Click Here To Show Diagram Code: [go]$$B diagram 2 $$ ----------------+ $$ . . . . . . . . | $$ . . . . . . . . | $$ . . . . . O . . | $$ . . X . X O . . | $$ . . . . . . . . | $$ . . . . . O . . | $$ . . . . . . . . | $$ . . . . . . . . |[/go]

These shapes are also related to ignoring the slide which you can make some tewari analysis with against 3-3 joseki:

Click Here To Show Diagram Code: [go]$$B $$ ----------------+ $$ . . . . . . . . | $$ . . . . . . . . | $$ . . 3 . . 6 . . | $$ . . . . 1 . 4 . | $$ . . . . . . . . | $$ . . . . . 2 . . | $$ . . . . . . . . |[/go]

1-3 is joseki (if 3-3 is an okay opening move), but then 4 is probably aji-keshi but 5 is a slack answer as blocking at a is stronger. Black a for b would then likely be a bad exchange.

Click Here To Show Diagram Code: [go]$$W $$ ----------------+ $$ . . . . . . . . | $$ . . . . . . b . | $$ . . 2 . . 1 . . | $$ . . . . 4 a 5 . | $$ . . . . . . . . | $$ . . . . . 3 . . | $$ . . . . . . . . |[/go]

P.S. Haylee just did a review of this game here:

Uberdude · Post by **Uberdude** » Mon Jan 09, 2017 7:25 am

Here's the Ke Jie vs Master game:

And Park Junghwan also making the connection:

And then Master as white answering Chen Yaoye's approaches and not taking gote to make the connection, Chen then blocked (and later Master did the attach on top of the Chinese opening stone to sabaki there as also seen in AlphaGo self-play games).

Another game with the same opening (rotated) as vs Park Junghwan above, but now Ke Jie is black and when he develops the moyo on move 13 Master dives in to make sabaki rather than connecting and allowing an even bigger moyo like Park did. It even ignores the block there later! So flexible.

Fan Tingyu making the connection, Master plays fast-paced opening. Later closes the side and then does the 2nd line poke at the thinness and ends up cutting the approach stone off!

pookpooi · Post by **pookpooi** » Mon Jan 09, 2017 11:17 am

so anyone can sum up v.25(rumored) style yet, from these 60 games...
As v.13, Fan Hui said it's calm and conservative while Ann Younggil described it as slightly territorial but well-balanced.
In v.18 Michael Redmond commented that it's more aggresive and unorthodox.

Uberdude · Post by **Uberdude** » Wed Jan 11, 2017 2:50 am

Uberdude wrote:Another of Master's favourite moves is the 2 space high corner enclosure from a 3-4, which is pretty rare to see in normal early-board situations in contemporary pro play. I wonder which pro will be the first to copy it? . Will it become a new fashion, like AlphaGo game 5 opening or that peep?

Well, I suppose Huang Yunsong did it before Mi Yuting in the Mingren semifinal yesterday!

Click Here To Show Diagram Code: [go]$$Bcm1 Huang Yunsong 6p (black) vs Mi Yuting 9p $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . 8 . . . . . | $$ | . . . 2 . . . . . , . . . . . 1 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 9 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . 7 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 6 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . 5 . , 3 . . | $$ | . . . 4 . . . . . . . 0 . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

Click Here To Show Diagram Code: [go]$$Bcm11 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . 9 . . . . . . . . | $$ | . . . . . . . 4 . 1 8 . . O . 5 . . . | $$ | . . . O . . . . . 7 6 . . . . X . . . | $$ | . . . . . . . . . 0 . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . X . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . 3 . . . . | $$ | . . . . . . . . . . . . 2 . . . . . . | $$ | . . . , . . . . . , . . . X . , X . . | $$ | . . . O . . . . . . . O . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

dfan · Post by **dfan** » Wed Jan 11, 2017 6:50 am

People following this thread now or in the future may like to know that there is another interesting discussion in progress of Master's style in the thread translation of a 9d pro commentary, starting with post #36.

Bill Spight · Post by **Bill Spight** » Fri Jan 13, 2017 11:45 am

For some time, strong Monte Carlo bots have played a more center oriented game than humans. Maybe because Go is a topological game, and stones in the center are more likely to connect up in Monte Carlo playouts.

AlphaGo does not seem as much center oriented, not even to the point of playing Cosmic Go, but it seems to me that it is more center oriented than today's pros.

To check that impression out, I came up with a very crude way to assess centrality. At the end of the game, how often does AlphaGo control tengen or occupy it with a living stone, by comparison with its pro opponent? By my count, in its 60 game streak it did so in 30 games, its human opponent did so in 15 games, and neither did in 15 games. That shows a decided preference for occupying or controlling tengen by the end of the game. However, since AlphaGo won all of the games, it has a decided preference for occupying or controlling any point on the board.

We really should compare the 2:1 odds of doing so for tengen with the odds of doing so on average for any point on the board, or for other selected points. If anyone wants to do that research, be my guest.

However, let us suppose that all the games were played out and scored. How often would AlphaGo control or occupy tengen? Conservatively, let us share out the 15 undecided games evenly. Then AlphaGo would occupy or control tengen in 37.5 games, on average. If we projected that probability onto every point on the board, AlphaGo would win by around 90 points. Nobody thinks AlphaGo is that good. So I feel pretty confident that AlphaGo garners tengen and other central points more often than its opponents.

Majordomo · Post by **Majordomo** » Fri Jan 13, 2017 12:38 pm

While probably wildly inaccurate - could some measure be made for the amount of points made on the 4th line or higher?

EDIT: As a ratio perhaps, to points in the centre vs points on the sides / corners, and how that ratio compares to the pros - less interesting with just an amount - since it always made more points than it's opponents overall

andrewgr · Post by **andrewgr** » Sat Jan 14, 2017 12:07 am

Bill Spight wrote:To check that impression out, I came up with a very crude way to assess centrality. At the end of the game, how often does AlphaGo control tengen or occupy it with a living stone, by comparison with its pro opponent?

Another very crude method, which would also potentially shed some light on your hypothesis about playing in the center due to ability to connect up, would be to count the number of living groups for each side at the end of the game. If your hypothesis is correct, I would expect Alphago to have fewer living groups, on average.

Bill Spight · Post by **Bill Spight** » Sat Jan 14, 2017 1:16 am

andrewgr wrote:
Bill Spight wrote:To check that impression out, I came up with a very crude way to assess centrality. At the end of the game, how often does AlphaGo control tengen or occupy it with a living stone, by comparison with its pro opponent?

Another very crude method, which would also potentially shed some light on your hypothesis about playing in the center due to ability to connect up, would be to count the number of living groups for each side at the end of the game. If your hypothesis is correct, I would expect Alphago to have fewer living groups, on average.

Well, AlphaGo, while it makes use of Monte Carlo evaluation, relies so much upon neural networks that I doubt if random connection plays much of a role in move selection.

Cassandra · Post by **Cassandra** » Sat Jan 14, 2017 3:15 am

andrewgr wrote:Another very crude method, which would also potentially shed some light on your hypothesis about playing in the center due to ability to connect up, would be to count the number of living groups for each side at the end of the game. If your hypothesis is correct, I would expect Alphago to have fewer living groups, on average.

Interesting idea.

When trying to find some "principles" in the courses of "P (B wins)" of CrazyStone's analysis' of the 60 games (the results of the "fast" analysis' can be found at viewtopic.php?p=215445#p215445), I got the "feeling" that the number of territories might play a role.

There are a few games (e.g. #9, #36, #48; not only one as stated in the posting linked above) that are -- according to CS's analysis -- more or less "undecided" until "late" in the game. In these games, the course of "P (B wins)" starts to raise significantly in AlphaGo's favour around moves 150 to 170. "Usually", this "point of no return" can be found around move 90 to 100 or even earlier.

I think that -- in these games -- the board is divided into a lot of (smaller) territories (at least more territories than on average).
Probably -- especially if these territories are not yet finally settled and / or are more or less "independant" from each other -- evaluation of the game is somewhat "more difficult" for AlphaGo.

AlphaGo's self-played game #3 (with longer time-setting) has this "a-lot-of-territories" feature, too, and this game has CS's "P (B wins)" between 0.55 and 0.45 troughout the entire game until the very end.
AlphaGo's self-played game #2 (short time-setting) shows the usual "point of no return" around move 100, does not have "so many" territories, but also some large ones.

Mike Novack · Post by **Mike Novack** » Sat Jan 14, 2017 7:47 am

Bill Spight wrote:For some time, strong Monte Carlo bots have played a more center oriented game than humans. Maybe because Go is a topological game, and stones in the center are more likely to connect up in Monte Carlo playouts. AlphaGo does not seem as much center oriented, not even to the point of playing Cosmic Go, but it seems to me that it is more center oriented than today's pros.

We should not ignore the possibility that this is the CORRECT way to play, but that it depends on being able to make use of widely scattered small amounts of aji << a large total amount of aji >> and that humans just can't manage that. That while collectively adding up to a lot of aji, the individual pieces of it are too difficult for the human to keep track of. Think of how these different little bits of aji migt be "discovered" by the multitude of lines in the MCTS playouts.

Calvin Clark · Post by **Calvin Clark** » Sun Jan 15, 2017 9:38 pm

Uberdude wrote:Another of Master's favourite moves is the 2 space high corner enclosure from a 3-4, which is pretty rare to see in normal early-board situations in contemporary pro play. I wonder which pro will be the first to copy it? . Will it become a new fashion, like AlphaGo game 5 opening or that peep?

He's not the first, but Xie He 9p tried it in the Tianyuan a couple of days ago:

Incidentally, few of AlphaGo's moves shock me. I think most of them were played by O Meien, who must be having a great laugh right about now.

John Fairbairn · Post by **John Fairbairn** » Mon Jan 16, 2017 4:18 am

Uberdude wrote:

Another of Master's favourite moves is the 2 space high corner enclosure from a 3-4, which is pretty rare to see in normal early-board situations in contemporary pro play. I wonder which pro will be the first to copy it? . Will it become a new fashion, like AlphaGo game 5 opening or that peep?

It depends, of course, on what you mean by 'pretty rare' or 'early' or 'contemporary' but there have been several pro games this year prior to AlphaGo's in which it appeared as early as move 5 (i.e. same as AplhaGo) and if we extend the search to move 10 or 20 we can get very many more examples - hundreds actually if we go back to the first player to 'copy' it, Honinbo Jowa, 200 years ago.

Since AlphaGo did not play this move until just after a brief flurry of human games with it, I'd be interested to know whether it is the copier, perhaps having some sort of weighting on recentness of training games so as to stay 'contemporary'.

The following opening was played just a fortnight before AlphaGo's first assay with the high two-space shimari. Otake Hideo played it on move 3. There were more normal examples just before.

Uberdude · Post by **Uberdude** » Mon Jan 16, 2017 4:59 am

By "pretty rare" I'd mean something like a few percent of shimaris, and if I look on ps.wlatheri for the next move in an empty quadrant from a 3-4 (so that includes mini-Chinese type moves too) the normal small low shimari is #1 with 32%, mini-Chinese #2 with 15%, #3 is small high with 8.5% and #4 big low with 7.2%. Big high is #11 with 0.5%. By "early" I mean first 10 moves or so when players are making normal fuseki moves. Let's say "contemporary" is last 5 years, seeing as pro fashions seem to change quite quickly. I wasn't claiming AlphaGo invented this move, but I expect Huang Yunsong, Mi Yuting and Xie He playing it recently is not in imitation of Otake but of AlphaGo. That Otake example is interesting, but I'm guessing white 2 was j15 so if that's the case the big high shimari actually makes more sense to counter centre potential than in a normal "players taking the corners on 3rd/4th line" opening. As an example of that here's Wang Zejin 4p (a player I hadn't heard of and #114 on goratings so presumably a young improving Chinese pro) doing it in July last year (full game: http://ps.waltheri.net/database/game/73362/). And O Meien and Gan Siyang have been doing it for ages, and I found Luo Xihe did it 5 times in 2015. It's like that early peep in the 3-4 high approach attach tiger mouth joseki that AlphaGo played, some small-name pro had played it a few years ago but no one takes much notice of them, but when the new super-strong player in town does it people pay more attention and start copying it.

Click Here To Show Diagram Code: [go]$$B Black: Wang Zejin 4p vs Jiang Weijie 9p 2016-07-04 $$ +---------------------------------------+ $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 2 . . . . . , . . . . . 1 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . 5 . , 3 . . | $$ | . . . 4 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ +---------------------------------------+[/go]

P.S. I just saw that the Nihon Kiin event on the Master games includes O Meien as one of the commentators. I wonder what he will say, "I told you this shimari was good!"

? https://www.nihonkiin.or.jp/event/area/ ... aster.html. Also Ohashi Hirofumi (the other commentator beside Michael Redmond) posted on his blog about a move from Kono Rin in the recent Kisei match that's a bad style one Master also played but I couldn't understand the machine translation well: http://blog.goo.ne.jp/minamijyuujisei_1 ... fdb16bba01.

John Fairbairn · Post by **John Fairbairn** » Mon Jan 16, 2017 5:41 am

It's like that early peep in the 3-4 high approach attach tiger mouth joseki that AlphaGo played, some small-name pro had played it a few years ago but no one takes much notice of them, but when the new super-strong player in town does it people pay more attention and start copying it.

While not disagreeing with the tenor of what you are saying - fashion following was the main driver behind New Fuseki - the tiger-mouth peep was not uncommon in the 1960s and was played by the likes of Fujisawa Hideyuki and Go Seigen. I suppose we may have to assume modern pros really are not familiar with the games of even Go Seigen.

Life In 19x19

Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master

Re: Learning from the Master