Top 10 Shocking Winning Percentages of the AlphaGo Teaching Tool

Definitely clickbait.

The AlphaGo Teaching Tool (https://alphagoteach.deepmind.com/) has had an ever-lasting impact on our canon of joseki and fuseki. It is certainly worth clicking through a few openings, even though most of what is offered is rather complicated.

Below is a summary of WeiqiTV‘s selection of the most note-worthy surprises from “AlphaGo教学工具之十大恐怖胜率 上集”. Note that the numbers in the diagrammes from the Teaching Tool represent Black’s winning probability in percentage points. I.e. Black should go for the moves with higher numbers, whereas White seeks the move with the lowest number.

#10)

att1

The most basic joseki, previously considered even.

This sequence apparently still gives an even game, but both sides have made mistakes: Black 11 and White 14.

What should be played instead:

att2att3

The most dramatic revelation in this joseki is that White shouldn’t connect after Black’s hanetsugi, which for decades has been a no-brainer. However, if White crawls instead, Black’s joseki choice proves to be slightly questionable.

Black can punish White’s 5.5%-mistake thusly:

att3aatt3b

#9)

att4

A former standard joseki.

According to AlphaGo, Black has already earned himself a losing position after this joseki.

I remember a tewari that claimed that White is better off by two points in this joseki:

att4a

In this even shape, Black makes an A-B exchange from inside, and White gets an outside stone in C-D, hence the two-point difference. That’s a disadvantage Black would actually be willing to endure as this sequence is happening in a corner White claimed first. But it seems that the aji of White E is worth much more than that.

Where did Black go wrong?

att5

Black apparently should have blocked. The usual move loses a terrifying chunk of 9.2% winning rate. The block leads to the following sequence:

att6

The tree splits up at this point. If you’re interested, click through the paths in the Teaching Tool. These variations haven’t been completely deciphered yet. Most likely this two-space high pincer will become the next taisha in terms of complexity.

#8)

att7

This one doesn’t concern me too much, as this fuseki hasn’t had time yet to migrate to Europe. The previous thought was that White should tenuki after 17. Letting Black take at 5 and connect at 10 can be tewarised to a shape in which Black has the ordinary shimari at 10+11 and White the extension at 14. This used to lead to complex judgement games of White aiming at the ko at A.

But the answer is simpler than that:

att8

White should just connect and Black is already on the losing side.

att9

In the sequence proposed by AlphaGo, the lower left shape becomes an even result. Humans would accept it only begrudgingly, as being first in a corner is supposed to give some sort of advantage. What wasn’t taken into account was that, 1) the position of the marked komoku is far inferior to a hoshi, letting White enter at A too easily, 2) Black’s pincer at B is positioned too closely to his thickness; C would be preferable, or at least Black could have stayed farther away with a pincer around D.

#7)

att10

Common opening in Mini-Chinese or Micro-Chinese/whatever you call it.

There are several mistakes in the lower right corner: Black’s entire strategy with 9, 11 and 13 (!), and White’s failure to take full advantage at 20.

att10aatt11

att13

WHAAAT?!

att13aatt14

White should push and pincer, so compared to the formerly standard sequence, Black does not get to extend majestically on the right.

att14a

But what the fish is this 😀 White doesn’t even tigermouth next move. Check it in the Tool.

#6)

att15

Another Mini-Chinese thing.

Apparently, there are countless examples of this happening in pro games. (I wouldn’t know as I don’t do Mini-Chineses.) But something has gone terribly wrong for White with the exchange of 16 for 17.

att16

White turning there instead of just retreating grants Black a whooping 12.7% in one blow.

att17

It also turns out that Black’s optimal response isn’t blocking, but the weird keima.

att18

In the mostly even sequence AlphaGo suggests, White takes a visibly larger chunk of the corner than before. Therefore, White is much worse off in the previous sequence.

att19

The reason for White exchanging the turn was discontentment with Black’s ability to block at 2. It turns out that the fight instigated by the cut 3 is actually good for White. Black absolutely needs to capture this stone whereas White can dodge and take advantage in any form whatsoever.

To be continued in part 2.

P.S.:

Here is an explanation of the Chinese video title because, as I found out last week, people crave that stuff.

AlphaGo 教學/教学 工具 恐怖 勝率/胜率
 ā fā gòu jiào xué gōng jù zhī shí kǒng bù shèng lǜ shàng
AlphaGo__ education__ tool_____ [poss.] ten big frightening win frequency__ upper episode

(之 zhī=”possessive particle, literary equivalent of 的 de.“)

In the interlinear transcription, we find that the title consists of a bunch of noun phrases strung together, followed by an ominous “upper episode”.

It’s a habit to suffix two of episodes, volumes, whatever of a series by 上 shàng “up” and 下 xià “down” instead of numbering them. When there are three of them, 中 zhōng “middle” is squeezed in.

I have no clue why this is done. I see little sense in stacking books vertically, and it seems impractical to tilt your head by 90° while browsing bookshelfs to visualise the top-down order. But thinking about this made me realise something:

When books are numbered normally, you may be missing several of the latter volumes without noticing that your series is incomplete. In the 上中下 order, this can only happen if you are missing the middle volume, and that one only. Perhaps this system is put in place to prevent major book loss. But it is only applicable when there are three of a series. I have no idea how a fourth one would be named under this system.

One thought on “Top 10 Shocking Winning Percentages of the AlphaGo Teaching Tool

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

w

Connecting to %s