What about this Alphago Joseki?

Gomoto · Post by **Gomoto** » Fri Sep 28, 2018 7:35 am

What do you think about this two positions?

: ajo1.jpg (164.22 KiB) Viewed 19913 times

: ajo2.jpg (165.45 KiB) Viewed 19913 times

Uberdude · Post by **Uberdude** » Fri Sep 28, 2018 8:19 am

Compare them in the same game

gowan · Post by **gowan** » Fri Sep 28, 2018 3:45 pm

For years I have objected to the common practice of using "joseki" to describe any and every sequence of moves. The advent of superhuman AI in go brings up a question as to the meaning of now joseki. Traditionally joseki refers to standard moves or sequences, the standard quality coming from history of use and consensus as to appropriateness of moves or sequences. Now it seems that simply being played by a top level AI go player makes something joseki. What do we make of this? Human judgment appears no longer to be needed.

Kirby · Post by **Kirby** » Fri Sep 28, 2018 4:44 pm

gowan wrote:Now it seems that simply being played by a top level AI go player makes something joseki. What do we make of this? Human judgment appears no longer to be needed.

After AlphaGo, a lot of people noted that they realized now how flexible go is - how many possibilities exist that they didn't think possible before. The "rules" of the past were no longer rules. In that spirit, I think it'd be a mistake to make new rules out of AI percentages. The value of top-level AI recommendations is that they bring about new ideas.

Human judgment is what we can use to make use of these new ideas.

AI should break down the barriers of rules we've made about the "proper way" to play - not make new rules.

sorin · Post by **sorin** » Fri Sep 28, 2018 9:57 pm

gowan wrote:For years I have objected to the common practice of using "joseki" to describe any and every sequence of moves. The advent of superhuman AI in go brings up a question as to the meaning of now joseki. Traditionally joseki refers to standard moves or sequences, the standard quality coming from history of use and consensus as to appropriateness of moves or sequences. Now it seems that simply being played by a top level AI go player makes something joseki.

I don't think the meaning changes much, as long as we let the word mean "commonly played sequence by strong players".
Whether the strong players are human pros, or AI programs.

Gomoto · Post by **Gomoto** » Fri Sep 28, 2018 10:38 pm

I am exactly interested in your human opinion on the two positions. What do you think of the possiblities, differences and characteristics of the two boards.

Elom · Post by **Elom** » Sat Sep 29, 2018 5:15 am

sorin wrote:
gowan wrote:For years I have objected to the common practice of using "joseki" to describe any and every sequence of moves. The advent of superhuman AI in go brings up a question as to the meaning of now joseki. Traditionally joseki refers to standard moves or sequences, the standard quality coming from history of use and consensus as to appropriateness of moves or sequences. Now it seems that simply being played by a top level AI go player makes something joseki.
I don't think the meaning changes much, as long as we let the word mean "commonly played sequence by strong players".
Whether the strong players are human pros, or AI programs.

Joseki = Phrase, one that hopefully uses good grammar and syntax
Shapes = Words, with Haengma being Verbs
Moves = Letters

Since joseki means 'set pattern', IMO the ones that fit the description of the word best are those from the endgame and middlegame

.

I don't know how helpful learning thousands of phrases of a language without understanding its grammar is, so I guess that's where the importance of understanding set patterns lie.

I find it interesting that in both pictures the right side is a big point, but there are other urgent moves to play; this leaves the one point jump as a valid but weak choice. However, in the first picture the pincer, if I may call it so, on the upper side scores best, yet the second picture favours a move that was took weak to be enumerated in the first.

Is this from black getting stronger? White getting weaker? I can't help but wonder on the cause.

gowan · Post by **gowan** » Sat Sep 29, 2018 10:43 am

Kirby wrote:
gowan wrote:Now it seems that simply being played by a top level AI go player makes something joseki. What do we make of this? Human judgment appears no longer to be needed.
After AlphaGo, a lot of people noted that they realized now how flexible go is - how many possibilities exist that they didn't think possible before. The "rules" of the past were no longer rules. In that spirit, I think it'd be a mistake to make new rules out of AI percentages. The value of top-level AI recommendations is that they bring about new ideas.

Human judgment is what we can use to make use of these new ideas.

AI should break down the barriers of rules we've made about the "proper way" to play - not make new rules.

My post above was in response to the use of joseki in the title of this thread. Unusual AI moves do indeed suggest new ideas but they aren't joseki (yet). Maybe we need nothing more than "interesting move" to describe them. Whether moves or sequences are joseki does, too, depend on the level of the player. There is that old maxim "There is no joseki for the Meijin" which means that the Meijin is so strong that everything is done on the fly rather than relying on standard moves. This might now apply to the top AI.

Bill Spight · Post by **Bill Spight** » Sat Sep 29, 2018 11:58 am

gowan wrote:Whether moves or sequences are joseki does, too, depend on the level of the player. There is that old maxim "There is no joseki for the Meijin" which means that the Meijin is so strong that everything is done on the fly rather than relying on standard moves. This might now apply to the top AI.

Quote from an 8 dan giving a class years ago in Kyoto:

"If I play it, it's joseki."

Gomoto · Post by **Gomoto** » Sat Sep 29, 2018 12:21 pm

My post above was in response to the use of joseki in the title of this thread. Unusual AI moves do indeed suggest new ideas but they aren't joseki (yet). Maybe we need nothing more than "interesting move" to describe them. Whether moves or sequences are joseki does, too, depend on the level of the player.

Just for the uninitiated, the sequence was played by Alphago, but there are also over 20 pro games with this joseki in my database

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

Uberdude · Post by **Uberdude** » Sat Sep 29, 2018 1:06 pm

gowan wrote:My post above was in response to the use of joseki in the title of this thread. Unusual AI moves do indeed suggest new ideas but they aren't joseki (yet). Maybe we need nothing more than "interesting move" to describe them.

How about just "AlphaGo sequence", though I think it's not far off joseki. This sequence is not a one-off that AlphaGo played once and we noticed, it is a sequence you will find many times in the AlphaGo teaching tool. So if joseki can be understood to mean "after many thousands or millions of games by a strong player (be it human or computer) it has become accepted as a good sequence" then it's joseki. Leela Zero and Elf have also both independently discovered it (actually recent LZ is not independent of Elf but is influenced by some Elf self-play inclusion) and come to the conclusion it is a good sequence for both sides in some positions. And pros have thought it worthy of trying out in several games; probably more pro examples than some of the more esoteric lines in your joseki books.

Gomoto wrote:I am exactly interested in your human opinion on the two positions. What do you think of the possibilities, differences and characteristics of the two boards.

First sequence with the falling-back hanging connection of 4 still feels a bit odd to me but here is some local analysis. White has not made the exchange of a-b compared to the second sequence. The plus for white is the stone at 6 has an extra liberty, the downside is the atari is not 100% sente so if you play it latter black might not connect (and then defending at 4 becomes basically redundant). One way this could happen is with the peep at c which is a key follow-up point: if white can play this and black defends at d that's nice for white to get the exchange whilst preserving the liberty, but maybe black would tenuki or resist (e.g. at e, also possible with 2nd sequence but there is a more immediate threat to cut).

Click Here To Show Diagram Code: [go]$$B $$ . . . . . . . . . $$ . . . . . . e . . $$ . . . . . . . c . $$ . . . . . 7 X d . $$ . . . . . 6 a 5 . $$ . . . . . O 1 b . $$ . . . O . . 2 3 . $$ . . . . . 4 . . . $$ . . . . . . . . . $$ -----------------[/go]

Another difference is with black's eyespace in the corner: he has more (and 4 in-a-row life) with the atari and connection whereas with the 2nd sequence when black pushes in the corner white can hane at 2-1 and it's not locally 2 eyes. Set against that in the 2nd sequence white does actually have a cutting point (though for now captured with a ladder) but that weakness can be a resource for black (e.g. the Ke Jie vs AG game 2 "joseki", or the f16 shape attack Kang Dongyun demonstrated).

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

A key difference with this sequence is that after white 6 black has the choice to either take the corner or the side, first sequence denies black the corner choice so white dictates the direction, but takes a slightly more cramped local shape to do so.

A related idea is this new AG/LZ sequence with extend not hanging for 4 and then the ugly black connect of 7. Previously I think people have thought the empty triangle with the initial approach stone is too ugly to consider this and only ko instead, but the bots say black is generally ok here and white also made some dubious moves if you tewari.

Click Here To Show Diagram Code: [go]$$B $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . X . . $$ . . . . . 5 7 . . $$ . . . . . O 1 6 . $$ . . . O . . 2 3 . $$ . . . . . . 4 8 . $$ . . . . . . . . . $$ -----------------[/go]

Tewari 1: black double hanes with 3 (all good moves so far) and white captures the stone with 4-6. Usually this is considered a mistake but occasionally it is ok (e.g. strong around and trying to kill all of black). Then black should cut at a and either ladder/squeeze but instead played 7. White should say thank you and extend to b but instead played some lame inside move at 8. Which was worse 7 or 8? How about 4? Black being able to atari at b gives that group a lot more strength so isn't as eyeless and heavy as the clumsy farmer hat would be otherwise.

Click Here To Show Diagram Code: [go]$$B $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . X . . $$ . . . . . 7 5 . . $$ . . . . b O 1 4 . $$ . . . O . a 2 3 . $$ . . . . . . 8 6 . $$ . . . . . . . . . $$ -----------------[/go]

sort of tewari 2: black starts with lame pole connection 3. 4 is sensible (and slightly different shape on side). 5 is not great but white should answer it but instead plays crap 6.

Click Here To Show Diagram Code: [go]$$B $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . X . . $$ . . . . . 5 3 . . $$ . . . . . O 1 . . $$ . . . O . . 2 4 . $$ . . . . . . 6 . . $$ . . . . . . . . . $$ -----------------[/go]

Another idea, from attach against 4-4 knight enclosure. 1 is reasonably sensible (actually with small knight because white can hane outside maybe not so great in isolated position?). 2 hane inside is ok, though outside might be stronger. 3-4 is good. 5 is ok, 6 is terrible, should extend (or tenuki or pincer), 7 is terrible (should atari or extend on side).

Click Here To Show Diagram Code: [go]$$B $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . . . . $$ . . . . . . 7 . . $$ . . . . . 5 3 . . $$ . . . . . O 1 . . $$ . . . O . . 2 4 . $$ . . . . . . 6 . . $$ . . . . . . . . . $$ -----------------[/go]

Gomoto · Post by **Gomoto** » Sat Sep 29, 2018 1:13 pm

Thanks a lot!

Bill Spight · Post by **Bill Spight** » Sat Sep 29, 2018 2:04 pm

Click Here To Show Diagram Code: [go]$$W $$ . . . . . . . . . | $$ . . . . . . . . . | $$ . . . . . . . . . | $$ . . . . . X X . . | $$ . . . . . O a X . | $$ . . . . . O X . . | $$ . . . O . . O X 3 | $$ . . . . . O . 1 . | $$ . . . . . . . . . | $$ -----------------[/go]

I think Uberdude covered this, but just to make sure.

One of the advantages of not playing the atari at "a" is that after

,

is atari.

Uberdude · Post by **Uberdude** » Sat Sep 29, 2018 7:13 pm

It's quite interesting to do tewari looking at the bot win % move-by-move. So using Elf v1 on the following full board position like the Gu Kang game of starting this joseki by black after 2 inside approaches in parallel 4-4 opening and with an order in which each pair of moves are about as bad as each other:

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

- Before 1 attach black is at 60%, remember Elf v1 thinks black is better (52%) to start, plus Elf thinks answering black's first approach with the patient knight move is a fairly big -8% mistake (2nd one minor mistake).
-

attach is a fine move (-1.5)%
-

hane only move.
-

pole connection is strange move not making good use of the approach stone, -45%
-

weird pull back should be at 5, -48%. So slightly worse than 3. In the end result it's easy to see how the connect of 3 ends up inefficient with the approach to make an empty triangle, but the inefficiency of this stone is harder to spot.
-

not an urgent place (so white should tenuki) but locally it's a plausible move and if white answers not a major change in the position (a slightly bad exchange for black because now white's next move of capturing one stone is better eyeshape and yose than had black not exchanged). -24%
-

obedient answer, -16%. So what I said was slightly bad exchange Elf thinks is net loss of 8% for black, but 8% for Elf is not so huge (pros make more often).
-

inefficient duff move. If white just answers at a it's a bad exchange for black who didn't get much pushing from behind and a net loss of 17%. But white can tenuki and then this is -43%. Elf actually thinks clamp at a is best move on whole board.
-

small place. -46%. So a small net loss for white in giving black the duff empty triangle and taking this yose. If answer with extend was -27%.

At the end black is around 57.5%, so -1% when attach was played. The normal order of this sequence is initially Elf's first choice, but then it decides it wants to play the ko as black.

So my conclusion from this is although 7 looks stupid with the empty triangle, you can think of the corner counter hane as tricking white into playing some equally crap move in the corner (that's not so obviously ugly) instead of answering 7 with extend so it ends up fairly even.

P.S. Using LZ #157 instead the story is pretty much the same but with different numbers (-45% becomes -9%).

Uberdude · Post by **Uberdude** » Mon Oct 01, 2018 12:51 am

This AlphaGo joseki featured in at least 2 of the Samsung cup round of 16 games today. Both times used by winning player, one expected so not much to do with this joseki but Xie Erhao beating Park Junghwan was a good result for him and his advantage came from fighting based on this joseki. Analysis at viewtopic.php?p=237359#p237359.

Life In 19x19

What about this Alphago Joseki?

What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?

Re: What about this Alphago Joseki?