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.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

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

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.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

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.

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

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.

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.

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.

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.

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.

Just to make sure we are on the same page, I meant this one:

Click Here To Show Diagram Code

[go]$$W
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . O X X . . . |
$$ , . . O . O , X . . |
$$ . . . . . . 1 . . . |
$$ . . . . . . . X . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |[/go]

Not:

Click Here To Show Diagram Code

[go]$$B
$$ --------------------+
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . O X X . . . |
$$ , . . O . O , X . . |
$$ . . . . 1 . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |
$$ . . . . . . . . . . |[/go]

I really hate myself that I miss this wonderful conference that happen just about an hour ago, hope it'll upload full talk later #DLD17 Demis Hassabis just told that AlphaGo will focus on...

https://twitter.com/danielequercia/stat ... 3629748224
https://twitter.com/dwingenter/status/8 ... 5350348800

That's great news. It makes sense that they would focus on using AlphaGo as a teaching agent since that is groundbreaking work that will eventually transfer to other fields. Still, it's great to know that the project won't be put on the shelf and will continue to benefit the go community and hopefully, by extension, humanity.

Me, too.

I wish Google would release logs that show for each move the expected win probabilities, and some other things like the policy network's idea of likely moves etc. That would give us something to consider without any fancy teaching tools.

An Dongxu 6p tries out the tenuki after push variation Uberdude mentioned earlier in this thread:



This is a pretty weird game. Look at the position at move 25:

Click Here To Show Diagram Code

[go]$$B A ponnuki in the corner is worth what?
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . O . . . |
$$ | . . . . . O . . . . . . . O . X O . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 1 . . . |
$$ | . . . O X . . . . , . . . . X O X . . |
$$ | . . . O O X . . . O . . . . . X . . . |
$$ | . . . O X . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

A very similar fuseki to one Zhou played against Shi Yue a few weeks ago: viewtopic.php?p=214957#p214957. It seems he was sufficiently impressed by AlphaGo's (and Lee Chang ho before etc) attachment into the Chinese formation that it's now his standard way to deal with it. In the game with Shi black didn't make the ponnuki and Zhou pulled the stone out. An's e2 hane and e4 atari are also moves that are perhaps influenced by AlphaGo which humans would often feel crude for playing but AlphaGo doesn't mind making those exchanges.

Now there are too many recent games to mention where some Master moves are being used. Interesting times.

A general feeling I have about AlphaGo is that AlphaGo is willing to play "aji-keshi" moves, probably because it has already decided on its fuseki strategy.

For example, pros were hesitant to play the aji-keshi tiger's mouth peep Uberdude mentions *early on* in the game, because it eliminates the option of playing a different way later on.

I feel like AlphaGo has already decided the global strategy and isn't, therefore, concerned about a play that pros would locally consider to be aji keshi.

_________________
be immersed

There's a lot of interesting games in the Tianyuan tournament with some AlphaGo/Master ideas, but first a rather fun move that surprised me but makes sense in hindsight:

Click Here To Show Diagram Code

[go]$$W White (Chen Xian 5p) to play against Gu Li 9p, Tianyuan round 3
$$ +---------------------------------------+
$$ | O . O . O . O . . . . . . . . . . . . |
$$ | X O O O X O O O . . . . . O . . . . . |
$$ | X X X O X O X X . . . O . O X . . . . |
$$ | . X O X X X O . . O . . . X . X . . . |
$$ | . . O X . X O . . . . . X . . . X . . |
$$ | . O O X X O O . . . . . . . . X . X . |
$$ | . O X X . X O . . . . . . . . . X O . |
$$ | . X O X X O O . . . . . . O X X O X O |
$$ | . O O . . X . . . . . . . . O O O X . |
$$ | 3 2 X X . . . . . , . . . . . , . X . |
$$ | . O O . . . . . . . . . . X O X O O . |
$$ | . . . . X . . . . . . . . . . O . O . |
$$ | . . . X O X O . . . . . . X . O O X . |
$$ | . . O O O O . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . O X . |
$$ | . . . O . . . . X , . . . . X , X O . |
$$ | . . . 4 . X . 1 . . . X . . . X . X . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Hint:

Answer:

The whole game was very interesting, with flexible (or would we call it inconsistent if he lost?) play from Gu Li with some early shoulder hits like his game with Master, and then tenukis, and a fun ko at the top left and funny trades on the right side.



The next round Gu lost to Zhou Ruiyang 9p who went all-out Master style with the big shimari and shoulder hit, tenuki the top right 3-3. Funnily enough despite all black's shoulder hits white ended up with a large centre territory, but black got more in the trades on the sides.