O Meien on AlphaGo

Uberdude · Post by **Uberdude** » Fri Nov 15, 2019 2:35 pm

Before we come to that, though, let us get back to O's initial reaction, or his "logic". He reasoned that if the aji at A had worked at once, Black would have played there at once. He didn't so a move there by White wasn't urgent.

If we assume* pulling out the 2 stones is too much with g10 group still weak then it is only after black has defended at 77 that it is a serious threat. So it is only on one prior move namely 78 that White might have defended, and one move 79 that black might have pulled them out. I can see White wanting to play 78 as in the game to make sense of the wall built on the previous moves: by separating the black corner White ensures his wall isn't an eyeless dame stick but attacks and makes it 2 dimensional with some eye potential, breaks a lot of black territory and sets up the famous invasion for later (which wouldn't have been so effective had Lee found the right time to peep at o9).

Also just because Lee made the mistake of not pulling them out for 79, AG only plays the board assuming it is playing against an opponent as strong as itself so defended because it thought it was the best move against Black's best moves. It doesn't do the human psychology (probably) bad habit of "my opponent didn't notice this good move so I won't defend against it but greedily grab other points in the hope I can come back later" strategy. I say probably because when deployed well this probably can win you more games. Indeed a bot trained to use such strategies could probably beat a top human by say 30 instead of 20 points.

* Which Elf apparently doesn't: if it thinks White should have defended there before 78 then presumably it also thinks black should pull them out before 79 and can handle g10 weak too.

Bill Spight · Post by **Bill Spight** » Fri Nov 15, 2019 4:26 pm

Uberdude wrote:
Before we come to that, though, let us get back to O's initial reaction, or his "logic". He reasoned that if the aji at A had worked at once, Black would have played there at once. He didn't so a move there by White wasn't urgent.
If we assume* pulling out the 2 stones is too much with g10 group still weak then it is only after black has defended at 77 that it is a serious threat.

{snip}

* Which Elf apparently doesn't: if it thinks White should have defended there before 78 then presumably it also thinks black should pull them out before 79 and can handle g10 weak too.

White played the nose attachment to Black's two stones in question with

. Containing them was on Elf's radar for

onward. Running out with them was on Elf's radar for

onward. It was only after

that running out became Elf's first choice for Black.

Uberdude · Post by **Uberdude** » Fri Nov 15, 2019 10:06 pm

Ah, thanks Bill. So it was on Elf's policy network for a while as a move worth investigating, but with some reading rejected it. It's nice that our bot overlords agree with basic human heuristics.

John Fairbairn · Post by **John Fairbairn** » Sat Nov 16, 2019 3:20 am

Although I am only part way into O Meien's book, and I do tend to focus on language more than most because of my background, I have to say I have been struck by his use of language.

I have already mentioned his Latinism (and he also cites Hannibal's march on Rome) and the fact that, so far at least, he eschews many normal go terms such as 'thick'. Even when he does use them he has a habit (here) of making them stand out in some way. For example, he uses words like 'heavy' in quotes, as if to say, "I'm not really talking about heavy in the way you normally understand but in a special AI sense" (and that particular case is very pertinent in Game 2 we are about to look at). Similarly, when he talked about nerai he highlighted it by waxing lyrical about how there's many a slip twixt cup and lip and many far-off nerai targets are not achieved because too much goes on in-between. I think we can see from Game 1 why he had a bit of a hang-up about nerai!

Another example that struck me, which he uses more than once, is the phrase 'standard theory' 標準理論. That's perfectly clear and many Japanese wouldn't bat an eyelid at it, but he does put it in quotes and eschews the routine go word for go theory, kiri　棋理. Again it's as if he is grasping for new ways to talk about new things, which is especially difficult when those new things look just like old things. Game 2 had several such difficulties.

There were several surprises for him and other pros, but the biggest surprise was that they came in a connected passage. In Game 1 it was possible to believe that AlphaGo was just a tactical monster that saw much further ahead than humans, the nerai being a perfect example. But in Game 2, the connectedness of AG's play looked deep (and novel) strategically.

This was O's first highlight. He said the pros had been looking forward to seeing how AG would play as Black, and for the first few moves they were very happy with what they saw. But Black 13 (triangled) wiped the smiles off their faces. (For me it brought a smile because it made me think AG was chiding Yi for choosing the strangulated Chinese in Game 1, saying: "See this is how it should be done!").

It was not the idea of tenuki that astonished O. He said it's perfectly reasonable in the case when the squared stone has not been played and Black can then treat his remaining two stones in that area as light. The problem for him was that the addition of the third stone makes that group heavy (or "heavy", to be precise) and so a move at A is now required in order to relieve that heaviness. This, he says, is "standard theory".

O says that he and his fellow pros initially came to the conclusion that the Black tiger's mouth shape was indeed heavy but not really attackable just yet, and that's why AG could tenuki. The subsequent moves, however, brought in the notion that heaviness was the wrong concept. This shape is more of a foundation stone.

O does not discuss White 14, but you can almost hear his eyes pop at Black 15 (below). This got an immediate name - the sugu nozoki or 'immediate peep'. And he says it too goes against "standard theory."

But he adds that this move was quickly adopted by top pros (interesting that he specified "top" pros

) and he gives the rationale: Black's loss in forcing White to connect against the peep is trivial compared to the gain, which is that White will connect directly here, whereas later on he might respond in different ways (e.g. with a kosumi). It is therefore "good timing." O doesn't explain why White must respond with solid connection now rather than varying, but presumably you have to be a "top" pro to understand that. It seems, though, the main reason it has not been played up to now is that it lowers the temperature (shitabi ni naru).

But once again the subsequent moves can cast a new light, AG continued with a now well known and popular joseki in the lower left.

O's interpretation of this sequence is interesting. I suspect most amateurs now play it because it's fashionable, it's joseki and they know some of the lines. But O believed AG was using it strategically for sabaki (without the quotes, so nothing too special there), and the idea was for Black to settle himself here as a way of alleviating any forthcoming attack on his stones to the right (i.e. forestalling any possibility of a splitting or karami attack).

All of that was O's Highlight 1 of Game 2. As his first long section he looks at two or three highlights in each of the five AG-Yi games. Highlight 2 in Game 2 is something special: the fake shoulder hit - the shoulder hit in thin air. This is real Harry Potter stuff! Professor Albus O Dumbledore has revealed what the standard theory of wizarding tells us. What will Dobby the Elf be able to tell us, with his insights into dark secrets from Malfoy Manor?

Bill Spight · Post by **Bill Spight** » Sat Nov 16, 2019 4:39 am

John Fairbairn wrote:Although I am only part way into O Meien's book, and I do tend to focus on language more than most because of my background, I have to say I have been struck by his use of language.

I have already mentioned his Latinism (and he also cites Hannibal's march on Rome) and the fact that, so far at least, he eschews many normal go terms such as 'thick'. Even when he does use them he has a habit (here) of making them stand out in some way. For example, he uses words like 'heavy' in quotes, as if to say, "I'm not really talking about heavy in the way you normally understand but in a special AI sense" (and that particular case is very pertinent in Game 2 we are about to look at). Similarly, when he talked about nerai he highlighted it by waxing lyrical about how there's many a slip twixt cup and lip and many far-off nerai targets are not achieved because too much goes on in-between. I think we can see from Game 1 why he had a bit of a hang-up about nerai!

Another example that struck me, which he uses more than once, is the phrase 'standard theory' 標準理論. That's perfectly clear and many Japanese wouldn't bat an eyelid at it, but he does put it in quotes and eschews the routine go word for go theory, kiri　棋理. Again it's as if he is grasping for new ways to talk about new things, which is especially difficult when those new things look just like old things. Game 2 had several such difficulties.

(Emphasis mine.)

Many thanks, John.

It seems as if O Meien is doing what I said we humans would do, modify old concepts and perhaps come up with some new ones. Gradus ad Parnassum. More power to him!

As for the rest of your note. Much for me to mull over over breakfast.

Bill Spight · Post by **Bill Spight** » Sat Nov 16, 2019 7:46 am

John Fairbairn wrote:

This was O's first highlight. He said the pros had been looking forward to seeing how AG would play as Black, and for the first few moves they were very happy with what they saw. But Black 13 (triangled) wiped the smiles off their faces. (For me it brought a smile because it made me think AG was chiding Yi for choosing the strangulated Chinese in Game 1, saying: "See this is how it should be done!").

Backing up a few moves, Elf has a small disagreement with AlphaGo of the game (AlphaGo-Li). Elf thinks that

is a minor error, losing 7%.

Click Here To Show Diagram Code: [go]$$Bcm9 Elf's variation for :b9: $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X , 0 . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 5 . 9 . . . . . . . . . . . . . | $$ | . . 4 8 . . . . . . . . . . . . 2 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . 7 . . . , . . . 1 X , O . . | $$ | . . . . 6 3 . . . . . . . X O O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

As almost always, Elf prefers the solid connection in the bottom right corner. AlphaGo-Li played the approach at 12. That may be because it was trained on human play, right? That approach before connecting on the right was popular at the time. Elf prefers to make the solid connection and then approach from the bottom side. After

Black plays the Go Seigen/AI shoulder blow at D-07. White solidifies his corner and then switches to the top left corner. I suppose that

-

violate standard theory.

John Fairbairn wrote:It was not the idea of tenuki that astonished O. He said it's perfectly reasonable in the case when the squared stone has not been played and Black can then treat his remaining two stones in that area as light. The problem for him was that the addition of the third stone makes that group heavy (or "heavy", to be precise) and so a move at A is now required in order to relieve that heaviness. This, he says, is "standard theory".

By the mid-20th century pros knew that the joseki in the bottom right was problematic for Black. (When I was learning go I read something that Segoe wrote about that.) The bots have taught us that the problem lies with the extension on the bottom side. Sometimes you make it, but often it can, and should, wait. Note that in Elf's variation Black does not make that extension.

John Fairbairn wrote:O says that he and his fellow pros initially came to the conclusion that the Black tiger's mouth shape was indeed heavy but not really attackable just yet, and that's why AG could tenuki. The subsequent moves, however, brought in the notion that heaviness was the wrong concept. {Emphasis mine, WLS.) This shape is more of a foundation stone.

O does not discuss White 14, but you can almost hear his eyes pop at Black 15 (below). This got an immediate name - the sugu nozoki or 'immediate peep'. And he says it too goes against "standard theory."

Game diagram for convenience.

Click Here To Show Diagram Code: [go]$$Bcm9 Moves 9 to 16 $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . 5 . . . . . . . a . . | $$ | . . X , . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . b . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 6 . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 1 . . . . . . . . . . . . . 4 . . | $$ | . . . . . . . . . . . . . . . 7 8 . . | $$ | . . . O . . . . . c . . 3 . X , O . . | $$ | . . . . . 2 . . . . . . . X O O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Elf regards

as a minor error and thinks that White should invade the top right corner at a (where else?

). Interestingly, Elf also regards the AlphaGo peep as a minor error, preferring the human play at b by 5%.

I like these disagreements between the bots. They remind us not to take what they say as gospel. Sic semper tyrannis.

John Fairbairn wrote:But {O} adds that this move was quickly adopted by top pros (interesting that he specified "top" pros ) and he gives the rationale: Black's loss in forcing White to connect against the peep is trivial compared to the gain, which is that White will connect directly here, whereas later on he might respond in different ways (e.g. with a kosumi). It is therefore "good timing." O doesn't explain why White must respond with solid connection now rather than varying, but presumably you have to be a "top" pro to understand that. It seems, though, the main reason it has not been played up to now is that it lowers the temperature (shitabi ni naru).

OK. Everyone at the time began to think of the AlphaGo peep as a marvelous play. Now that we have more experience with bots, I think that we can take it as an example of a bot playing kikashi early, without worrying about aji keshi or the loss of a ko threat. Why the bots do that we do not know. Maybe it helps with reading by simplifying the position early. Anyway, Elf regards it as a minor error and prefers the human play at b. After which Elf leaves the peep unplayed. It does not appear in any variation after b, even though White immediately replies with the pincer at c.

So we have to question the timing of the peep. O Meien's reasons for playing it or not playing it are open to question. First, let me note that lowering the temperature (下火) seems to accord with the informal Western go term of temperature. 火 = fire, heat. But if so, why would Black not want the temperature to drop on the bottom side? (Aside from considerations of aji keshi and ko threats.) A lower temperature means that White would gain less from attacking Black. Second, as for playing the peep when White has to reply with the solid connection, Elf disagrees.

In fact, Elf regards

as a substantial 10½% mistake.

Click Here To Show Diagram Code: [go]$$Bcm15 One good peep deserves another $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . X . . . . . . . . . . | $$ | . . X , . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . 4 . | $$ | . . X . . . . . . . . . . . . . O . . | $$ | . . . . . . . . . . . . . 2 . 1 . . . | $$ | . . . O . . . . . , . . X 3 X , O . . | $$ | . . . . . O . . . . . . . X O O . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

takes advantage of the weakness of the hanging connection to peep at its mouth. Then

plays the second line kosumi.

So Elf casts doubt on the preferences of AlphaGo-Li. OC, we cannot take Elf's choices and winrate estimates as gospel, either.

Uberdude · Post by **Uberdude** » Sat Nov 16, 2019 11:27 am

Just a quick comment on the early peep of game 2. Later versions of AlphaGo, such as used for the teaching tool opening book, thinks it's a bad move if White replies correctly with the counter peep instead of submissively connecting as Lee did (but the old "standard theory" was that that wasn't submissive but made the peep aji keshi). Also earlier versions of AlphaGo tended to like that hanging connection, but later versions prefer the solid like Elf does, because they realised hanging connection is gote whereas solid is sente because it threatens the severe follow up of Go Seigen's attach and 2nd line hane. I think LeelaZero went through a similar evolution.

Gomoto · Post by **Gomoto** » Sat Nov 16, 2019 12:14 pm

Go Seigen's attach and 2nd line hane

Could anybody point me to a Go Seigen game with this play please?

Uberdude · Post by **Uberdude** » Sat Nov 16, 2019 12:30 pm

I don't know if Go actually played it, but afaik it was an idea he recommended in study groups or lectures and books he produced in later life after he'd retired from competitive play.

Gomoto · Post by **Gomoto** » Sat Nov 16, 2019 12:53 pm

Thank you, I did search for it in my database and did not find an example. That is why.

John Fairbairn · Post by **John Fairbairn** » Sat Nov 16, 2019 1:20 pm

Since various people are citing opinions based on different times (in a fast-noving field) and different machines, or even different versions of the same machine - all good stuff - it occurs to me to repeat that O's book is dated late 2017, i.e. ancient. He doesn't even mention Master (or Magist), or Zero anything, though he does spend a lot of time on DeepZen.

There is also the (maybe debatable) point that the earlier writers seem to have tried to anthropomorphise the bots. I have pointed out a couple of examples by O. The Korean 9-dan Hong Min-p'yo (with computer expert Kim Chin-ho) wrote a book on the same match in which Hong described move 80 in Game 1 as a "declaration of victory" by AG. In Game 2, he commented on the tenuki-ed Black shape we are currently discussing as demonstrating that AG "likes to build moyos rather than defend its groups". (And for Hikaru lovers, move 102 in Game 1 was a true kami no itte for him.)

I think we have all, pros and amateurs, moved on quite a long way in a very short time from that old use of terminology when talking about bots, and numbers now punctuate most articles and posts. But we are all still cleaving to certain words, and I find it fascinating to note which words seem to be retained and which are (apparently) being discarded. The words being retained are not necessarily those that were big in pre-bot days. Overconcentration has moved right up the charts; efficiency, too. Technical terms for moves such as peep and shoulder hit have Oscar nominations. Thickness now seems to belong to silent movies age of go rather than the dynamic talkies of today. Aji is hanging in there, but is maybe being overtaken by nerai. Some such a moyo have lost appeal for specific authors, while others keep them (but to me don't seem either convinced or convincing).

What specific examples do you see for this change in terminology? When you write about bots, which words do you now suppress? Which ones do you now highlight?

Knotwilg · Post by **Knotwilg** » Sat Nov 16, 2019 1:35 pm

John Fairbairn wrote: When you write about bots, which words do you now suppress? Which ones do you now highlight?

Nuance: I still don't write about bots. I think about Go, fueled by bot analysis. I don't think I suppress any words due to that: the balance between influence and territory has shifted a little to territory, but we do see examples where cooperating influence is favored over enclosed territory. Thickness is a term I try to avoid, not because of bots, but because of the discussions we've had about it.

Popular words/concepts: sente, tenuki, 3-3 invasion, peep, shoulder hit
New words: probability, policy, win rate,
Old/less popular words: honte, miai, split, probe

Gomoto · Post by **Gomoto** » Sat Nov 16, 2019 2:43 pm

highlights: shape, mistake, good move

SoDesuNe · Post by **SoDesuNe** » Sat Nov 16, 2019 2:53 pm

John Fairbairn wrote:What specific examples do you see for this change in terminology? When you write about bots, which words do you now suppress? Which ones do you now highlight?

Highlight: Efficency, overconcentration, attachments, tenuki.

RobertJasiek · Post by **RobertJasiek** » Sun Nov 17, 2019 12:41 am

AI play neither deemphasises thickness (or other concepts) nor raise importance of efficiency (or territory or other concepts). From AI play, some learn their own previous under-estimation of some concepts. Do not use that as an excuse to now neglect other concepts.

Means of positional judgement include
1) reading (or, for AI, emulated reading or a combination of both),
2) mostly static aspects (current territory, endgame-like counts...),
3) dynamic aspects (influence, fights, options...),
4) efficiency,
5) probability.

For reading, basic theory is known but excecution requires effort, time or (for AI) calculation power.

For static and dynamic aspects, there are well developed theories (especially mine about positional judgement and endgame evaluation). Their application often also requires (simplified) reading.

For human players, probability relies on experience and guesswork so is very unreliable. For AI, probability relies on sampling; due to its huge amounts, it is reliable for the strongest AI engines but has gaps (ladders, status assessment...) for below-strongest engines.

Quite a few professional players use all means of positional judgement. Not necessarily, all their usage of all means is good but some means they should use well for their own play. Mostly they hardly can explain their use.

This results in relative frequency of very short, easily generated comments, such as good versus bad efficiency. Now, we must not deceive ourselves by thinking that efficiency would be the only useful means, and very selective good versus bad assessments would be sufficient while almost all aspects and stones of positions might be considered ok without analysis and verification. Only comparing Black's and White's obviously most inefficient mistakes hones our laziness of not doing careful positional judgement.

It would be useful to develop good/bad efficiency from its current needles of occasional insight in a haystack to a well-developed theory, like for the (other) static and dynamic aspects. Currently, I do not see such a general theory of efficiency in terms of simple good/bad comparisons (very far beyond the only specialised tewari), unless we substitute the former as a tag for the theories of the (other) static and dynamic aspects. (Note that the same stones can have aspects of static and dynamic efficiency when, e.g., protecting territory and building influence.)

Rather than an aspect of efficiency, I consider impact of stones on fighting potential as part of the dynamic aspects. For that, I have already provided influence stone difference and excess counts of new valuable versus neutral/dead stones as very useful concepts. For identifying fights, I have introduced the concept of fighting value.

We perceive positional judgement by the strongest AI as strong because of excellent (possibly emulated) reading and probability-sampling.

Life In 19x19

O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo

Re: O Meien on AlphaGo