I found this interesting, I thought Master didn't like those high side pincers too much, I don't understand why it chooses this, it's because right and left side value is too settled by white 10 and 12? Still I would expect F17 and just let white double approach

I've noticed that AlphaGo seems to operate very much like Mizokami's stone-counting theory. Superficially, at least, that's not surprise because we can make the assumption that counting something or other must be a major part of any computer player.

Applying it to the above position, the first step is to decide which half of the board to concentrate on. There are four halves, of course. In the lower half Black has 7 stones to White's 8, so he can't take an attacking posture there. But his weaknesses are not yet glaring enough to require urgent remedy, so this side can be left for the moment.

The left side is 3-2 in Black's favour but that allows on a mild attack and Black actually has nothing to attack there, so that side is very ignorable at present.

On the right side Black is on the wrong side of a 6-7 split and so is required to play a prudent move - R14 springs to mind. That can hardly be bad but perhaps runs the risk of being slack or forced (with weaknesses on the lower side, it seems foolish for Black to surrender the initiative to White so easily - and a White reply at K16 makes the left side 3-3 which seems to dull Black's prospects a great deal on that side).

On the top side Black has 2-1 split and so by M's theory can allow himself a modest attack. The kick at P17 is too aggressive - no back-up. F17 is not even an attack and is likewise initiative-surrendering. That leaves a pincer. Given the uncertain situation on the lower side a high pincer seems obvious. The only two modest attacks are K16 and L16, both of which make the right side a safe 7-7. K16 is the more modest attack but also creates a 4-2 situation on the almost virgin left side, which gives Black something to aspire to.

So by my reading Mizokami's method predicts K16, and although it took a while for me to type out, it can be found well within the 30 seconds he claims is all you need to evaluate positions. And even if the method's not quite as accurate as AlphaGo, it does, by my reckoning, tend towards finding the same sort of moves, and perhaps for an underlying similarly numerical reason.

Alpha seems to agree with most humans that the 4-4 and 3-4 points in the corners are key. So I decided to look at the openings where each side takes two corners at 3-4 or 4-4 in the first four moves and see what the extremes are. In other words, how badly can either side go wrong if they do this at random? The range is pretty small, all I've found are between 45% and 48% for B. Alpha's estimates aren't that precise here - almost every time I find a position where the win rate is near one of the extremes after four moves, when I play out Alpha's suggestions to the end the win rate is back right near the average or even swings to the opposite extreme. Two examples: the first starts off about as well for black as I could find, at 48%, but the continuation ends with B's win rate for at 46.9%, which is nothing special.



And here's an even more dramatic shift, which starts off with B's win rate at 47.4% and ends with B's win rate down to around 45%

What does this say about the infamous " lost the game"? Here it's actually and !

Looking through AlphaGo's judgements of the Master games I came across a shocking judgement in a common joseki where it prefers (by a lot) a crude bumping itself on head to the "good style" normal joseki. Here's the sequence from Master vs Shi Yue (a rare occurrence of Master as black playing the small shimari for an orthodox opening), note that before 1 black's win% as reported in the teaching tool was 44.6, a bit of a drop from the 47 at start (most of this seems to be Master/Aja not playing the e5 peep before pressing at top left, I think it wants to break ladder before continuing the joseki, I might analyse more in the Master review thread later). This right side pattern is very common in pro games over the last few years, Shi then continued with kick of 5 and Master ignored because if you extend white can pincer which is also a ladder breaker.

Click Here To Show Diagram Code

[go]$$Wc AlphaGo Master (black) vs Shi Yue
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . X . . O . . |
$$ | . . . O . . . . . , . . . . . X . O . |
$$ | . . . . . 6 . . . . . . . . . . X O . |
$$ | . . O . . . . . . . . . . . . X O 3 . |
$$ | . . . . . . . . . . . . . . 4 1 2 . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , X . . |
$$ | . . . . 5 X . . . . . . . . X . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Instead of the hane of 1 above, which AlphaGo says is a huge mistake and launches white's win% by 8.7 up to 53.3, it wants to play this crude bumping and walking into double hane sequence (that low dans on Tygem seem to like, maybe they aren't so noob!), though it does interpose a cheeky peep, and then tenuki! if the tenuki is to b then it expects black to capture the 5 stone immediately, with a the area is left unplayed for the next dozen moves they give us.

Click Here To Show Diagram Code

[go]$$Wc AG's recommendation
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . X . . O . . |
$$ | . . . O . . . . . , . . . . . X . O . |
$$ | . . . . . . . . . . . . . . 3 4 X O . |
$$ | . . O . . . . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . 2 1 . . |
$$ | . . . . . . . . . . . . . . . . X 5 . |
$$ | . . . . . . . . . . . . . . . . . 6 . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . a . , X . . |
$$ | . . . . b X . . . . . . . . X . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I then checked if this was a specific recommendation in this whole board position with the two approaches already played on the left, but it's not. It also says it with just 2 white 4-4s, here the hane boosts black by 7.4 to 52%. I wonder if it's the lack of a severe ladder breaker for white that makes it dislike the hane so much.

Click Here To Show Diagram Code

[go]$$Wc AG's recommendation, plain double 4-4
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . X . . O . . |
$$ | . . . O . . . . . , . . . . . X . O . |
$$ | . . . . . . . . . . . . . . 3 4 X O . |
$$ | . . . . . . . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . 2 1 . . |
$$ | . . . . . . . . . . . . . . . . X 5 . |
$$ | . . . . . . . . . . . . . . . . . 6 . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . , X . . |
$$ | . . . . . . . . . . . . . . X . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Same thing with bottom left as 3-4, here its 7.7 difference: 51.5 vs 43.8.

Click Here To Show Diagram Code

[go]$$Wc AG's recommendation, plain 4-4 and 3-4
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . X . . O . . |
$$ | . . . O . . . . . , . . . . . X . O . |
$$ | . . . . . . . . . . . . . . 3 4 X O . |
$$ | . . . . . . . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . 2 1 . . |
$$ | . . . . . . . . . . . . . . . . X 5 . |
$$ | . . . . . . . . . . . . . . . . . 6 . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , X . . |
$$ | . . . O . . . . . . . . . . X . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

So, any bets of the first pro to open themselves to ridicule and play this crude move just because AG said so? There was the similar crude bumping in the game vs Gu Li, which Kono Rin then copied.

Wow that is pretty interesting. I don't quite believe it...

That seems to be quite a drop off in winning percentage for black. The teaching tool advised a 3-3 invasion for black at move 9 and at move 11 leading to better winning percentage > 48%. If I remember correctly, those early 3-3 plays weren't seen during the Master series but only later when the selfplay games were published (after the Ke Jie match). It seems Master played in a more "orthodox" way then.

_________________
Couch Potato - I'm just watchin'!

Yup, this teaching tool and the 50 self-play after Wuzhen are more 3-3 obsessed than Master at 2017 new year. There were a few earlyish 3-3s in the Master games (I'm planning to review the Gu Zihao one), but they were more like at move 29 than move 9. So I suspect although they say the teaching tool and 50 self-plays are AG Master version it's more like that version of the software design, but with more training so closer to AG's ultimate view of best play which is highly 3-3 obsessed. Another theory is when Aja was playing he had several move suggestions with win % from AG and chose not to go so 3-3 obsessed or play a small shimari for a change. Also Demis said they turned up the temperature parameter for the Master games so that would explain more variety of other moves not in the AG teach book, but not the total disappearance of the #1 choice move 10 3-3 invasions.

I think this 3-3 obsession is a bit of a shame really in the teaching tool, as so many positions its recommended mainline is just 15 moves of 3-3 invasions, which is not so interesting after the 100th time, and really it's the play after those sequences showing how the walls aren't so valuable that would be more instructive.

Kinda funny to see those old "greedyGo" BQM

https://senseis.xmp.net/?BQM83

https://senseis.xmp.net/?BQM133

When I was learning to play go, I wondered about , but came away with the idea that it was the pro choice, as opposed to simply connecting at , even when the ladder works.



I think that the "cheeky" peep is significant. It is a no loss kikashi. It induces bad shape by Black and does look better than the stone in the ladder in the other variation. I am surprised by how bad AlphaGo thinks that the hane is by comparison, but the peep is attractive. I am also surprised that AlphaGo prefers in this diagram to the solid connection at "c".

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

Visualize whirled peas.

Everything with love. Stay safe.

I agree. I kind of wish that the AlphaGo team had more go players on it.

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

Visualize whirled peas.

Everything with love. Stay safe.

As I understand it, top chess players don't take much notice of chess computers when trying to improve. Of course the computers are much, much stronger - by very much more than AlphaGo lords it over human go players - but that is of no practical significance for the human chess experts. I think it may have been Aronian who said they are just useful as blunder checkers and for preparing openings.

That situation may well apply in go, and with knobs on if AlphaGo really isn't a huge amount stronger than the best humans. There is, however, a possibly crucial difference, as revealed in this quote:



If AlphaZero did learn chess in a human-like way (and, by extension, likewise with go) that might be easier to tap into than the almost purely calculated mode of play in Stockfish et al. But at the moment, "human-like" seems to mean just a simulation of intuition, and so we already know how that is developed - by playing a massive number of games. Current experience seems to suggest the only way to get better tuition is to play even more games, which can only be done by starting a heavy regime at a younger and younger age. But surely we are nearing the limits for that now, and I can't imagine many parents would want their children to spend so much time on a game for which computers may well make professional events otiose.

Unless of course you believe you can learn go in the womb, as Shusaku's mother supposedly claimed.

I asked a pro about this recently, and he said his study group had decided that this White really is better than the old joseki. He mentioned getting the peep specifically as an important difference.

Something else that strikes me about this new AlphaGo joseki is if black goes all-in for some big moyo on the lower right quadrant without taking the one white stone, white might even capture the outside stone and sacrifice 2, though then q18 locally kills as an L+1 group but with some bad aji. Probably not as early as below as there were still less damaging ways to go inside, but that's the idea. Black could prevent this in sente by making the s13 s14 exchange and then tenuki, but it's a lossy exchange if white gets the next move there (-2 points endgame plus giving eyes).

Click Here To Show Diagram Code

[go]$$Wcm7 Oh you wanted that moyo did you?
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . 3 . . . . . . . X . . O . . |
$$ | . . . O . . . . . , . . . . . X . O . |
$$ | . . . . . . . . . . . . . . O X X O . |
$$ | . . . . . . . . . . . . . . . X O 0 . |
$$ | . . . . . . . . . . . . . . . X O 8 . |
$$ | . . . . . . . . . . . . . . . 6 X O . |
$$ | . . . . . . . . . . . . . . . . 5 X 9 |
$$ | . . . , . . . . . , . . . . . , . 7 . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . 4 . . . . . |
$$ | . . . 1 . . . . . . . . . . . . . . . |
$$ | . . . , . . . . 2 , . . . . . , X . . |
$$ | . . . O . . . . . . . . . . X . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I took a look at a few of the alphazero chess games, and there is probably a lot that chess players can learn now as the old chess engine design don't compare to alphazero. The old chess computers were based on lots of calculation, with standard estimates of queen=9 pawns etc. But in these alphazero often wins by sacrificing material and a very strong positional judgement (more human techniques), often putting Stockfish in a very long term bind and using zugzwang to make Stockfish make a bad move. Stockfish is unable to realise the trouble it's in since the compensation is so long term.

A series of comments on uberdude's diagrams

Click Here To Show Diagram Code

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

Note that alphago did play the above kosumi against Fan Hui I think

Click Here To Show Diagram Code

[go]$$Wcm16
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . 4 . . . . . . . . . . . . . . |
$$ | . . . X . b . . . . . . . . . . . . . |
$$ | . . . , . . 3 . . , . . . . . X . . . |
$$ | . . . 1 a . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 2 . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . O O X X X X X . , . . . . . O . . . |
$$ | . . . O O O O . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

That does look so bizarre, and will take me a while to understand before I use it. The first instinct is that it is so submissive that it can't be good for B (but I suppose that is the fault of the pincer). But I suppose it is no more submissive than many more familiar kosumis. But it still feels particularly bizarre and doesn't feature in my intuition at all. I guess I'm just overly influenced by standard joseki shapes.

Click Here To Show Diagram Code

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

But it did remind me of a Lee Changho Kong Jie LG cup final game which maybe demonstrates why keeping a solid hold on the corner is important: https://www.go4go.net/go/games/sgfview/23079

Click Here To Show Diagram Code

[go]$$Bcm5
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . 7 . . . . . . . . . . . . . . . . . |
$$ | . . . 6 . . . . . . . . . . . . . . . |
$$ | . . 5 O . . . . . , . . . . . O . . . |
$$ | . . 3 4 . 1 . 2 . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

I struggle to understand this one: without the 5-6 exchange, it is a standard shape, but how can the 5-6 exchange be good for B? Given that alphago rarely pincers, surely this is meant to be good for B? My previous understanding of the narrow pincer is that B can easily tenuki and W can't get much profit here. But I think I need to update that judgement. I may remain confused until someone else can explain how to treat these pincer variations.

Click Here To Show Diagram Code

[go]$$Bcm5
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . 6 . . . . . . . . . . . . . . . |
$$ | . . 5 O . . . . 2 , . . . . . O . . . |
$$ | . . 3 4 . 1 . . . . . . . . . . . . . |
$$ | . . 9 7 8 . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

also note that alphago tends to tenuki the above at this point. Crawling on the 2nd line would induce a jump, and connecting leaves very bad aji, so ...

Click Here To Show Diagram Code

[go]$$Wc AG's recommendation
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . X . . O . . |
$$ | . . . O . . . . . , . . . . . X . O . |
$$ | . . . . . . . . . . . . . . 3 4 X O . |
$$ | . . O . . . . . . . . . . . . X O . . |
$$ | . . . . . . . . . . . . . . . 2 1 . . |
$$ | . . . . . . . . . . . . . . . . X 5 . |
$$ | . . . . . . . . . . . . . . . . . 6 . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . a . , X . . |
$$ | . . . . b X . . . . . . . . X . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

This doesn't feel strange to me anymore. I use the bump in one of the 3-3 joseki quite a lot. My analysis of these bumps (including tewari, comparisons) that locally the efficiency is best. But that for a handicap game when you would need to invade the other side of the wall more deeply, then you shouldn't bump so crudely, and should play submissive locally (such as just connect), but aim at enemy weaknesses. I would be very tempted to use the bump in this joseki, though I'd be nervous that the right side and centre is pretty wide here. Though I don't know if I'll play the knight's move into the corner rather than attaching anytime soon.



There is that point, but I think it was nearly impossible pre-alphago to judge and analyse positions as accurately as we can now, so we had to just keep exploring variations (the magic of the unknown) and follow in other's footsteps. I would often try to use logic/tewari/direction of play to analyse openings but gave up around 1d, following the standard pro saying "the opening is irrelevant, you can play anything and it doesn't matter" (i.e. it isn't worth analysing it). There are so many different openings/variations, and I would try to analyse them, but couldn't find solid logic to base conclusions on, and there are too few games to base conclusions around statistics.

Back then, when deciding a move, consistency and coherence of a plan (frequently fighting spirit) was emphasised far more than judgement, or whether one move was known to be better than another.

Only post-alphago, when we have so many "correct answers" to compare to, do I feel we can much more confidently and accurately judge positions and explain clearly why certain moves are better than others, and hence more importantly how to use them. And it isn't that we are just listening to alphago's moves, but 1) we can extrapolate from correct answers, intuitively finding the threads and themes that link good moves, and analysing where the most persistent flaws in our own thinking lie. and 2) use logic/analysis on that intuition to work out how to follow up play. (I feel like my human brain is a good neural network that trains well on "correct moves" in certain positions and can extrapolate, just as the early alphago learnt off human moves).



In my opinion, you shouldn't read too much into the fine details of winning percentages/only move/logical relationships that are implied by the alphago numbers. I used to overvalue alphago's answers, but now I don't think alphago is close enough to perfect for some of this to be accurate. We can easily rely on something if it says one move is better than another by 3%, but inferring things from anything closer seems risky, let alone subtle/precise concepts such as these logical relationships. I also don't know how long alphago was allowed to think to create these numbers, what positions were not included, etc. And deeper down the tree, presumably the numbers get less and less accurate.

I copied a long list of things down on the day the tool was released. I guess there's no reason not to share it, as we seem to have picked up on very different things:

Click Here To Show Diagram Code

[go]$$Bcm1
$$ | . . . 6 . . . . . ,
$$ | . . . . . . . . . .
$$ | . . 4 . . . . . . .
$$ | . . @ 3 5 . . . . .
$$ | . X 1 2 . . . . . .
$$ | . . . O . . . . . .
$$ | . . X O . . . . . ,
$$ | . . X O . X . O . .
$$ | . . . 7 . . . . . .
$$ | . . . . . . . . . .
$$ +--------------------[/go]

I tried this (the B6 move recommended by alphagozero) recently against Zen7 set at 3d strength. It responded immediately with the marked stone (interestingly, when i analyzed the game thereafter with zen, ite did not even consider the marked move as one of the possible choices). The push and cut sequence to 6 looks pretty natural.

What i found interesting was move 7. Despite showing initial preference for f5 and h4, Zen7's preferred move (after 80,000 iterations) was 7, which is back to the hane once again. This seems odd to me, since i thought the whole point of sliding to B6 was to avoid making this exchange and damaging the F3 stone. Zen's earlier preferred moves of h4 seem more consistent with the B6 plan, in my view. Any thoughts on this?

This corner sequence is very interesting to me. Thanks to all here who pointed it out and discussed it. I have not been able to find any alphagozero sequence where white continues in the corner after B6 (usually W tenukis and then B eventually plays at the marked position). Has anyone else managed to find such a sequence?

Push and cut doesn't look that natural to me due to b selfharming the group, and W having a base on both sides. I think bots have taught us we can tenuki the marked move and that crawling on the second line can actually be acceptable if W takes D2. Probably alphago would hane and connect and call the marked move slow for being gote. This particular cut resembles a joseki where b has approached from the other side where the cut is premature and the f3 h3 exchange seems questionable for B.

There really isn't much negative to 'damaging' the F3 stone if W responds by connecting solidly as humans do, as the stone still occupies a vital point damaging w's shape. I think the second line keima does preserve sabaki aji if B gets the next move locally such as with h4, but running heavily to counterattack with the f3 stone still seems very unrealistic as d2 and e2 have such an impact on the corner and f3 is already so tightly pincered). So the 2nd line move is mostly to make it hard for w to find a follow up locally, but you need a bit of a plan on how to follow up yourself, for which perhaps we need to ask alphago. Otherwise, just play the human joseki. In this case, I see nothing wrong with it.

Edit: actually how to follow up for b seems obvious, b will defend at c7 or d7 so d2 and e2 are less of a threat and ask W how to defend. And wherever w defends, aji will remain, either b will still have the chance to hane and connect, to run with f3 or use the j3 aji. The complication is that W doesn't have to defend immediately though B can fight strongly with f2 (nb w can still tenuki). But b will have taken a healthy profit locally this way.

But if W follows up first such as with Zen's move, b hane and connect and somehow B got sente in a joseki where w normally expects sente.

Still, the human joseki settles things locally and is attractively simple. I would have to consider in a serious game whether to play the alphago keima as it is initially submissive, so I'd still need to be alert for plans to do something with f3, and plenty of reading/risky fighting could ensue. On a bad day that might just be making trouble for myself.

Regarding your final question, I remember finding an alphago variation where w immediately played d8 given a wall above from the 3-3 invasion of the star point. And b immediately played hane and connect. I think it was from this one space low pincer but I don't remember if W then tenukied or continued

Click Here To Show Diagram Code

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

In this shape also, AlphaGo thinks the crude bump at A is quite a bit better (about 3-5 percentage points) than the normal move at B, regardless of the global position. I guess it's an attempt to get sente, although quite often AlphaGo will simply tenuki the second line hane like in the game you linked, but if that happens, then white has gained the option of playing either the third line hane/atari or the cut, while in the normal joseki, cut is the only good follow up for white. AlphaGo loves having options.