So what's wrong with sanrensei?

Bill Spight · Post by **Bill Spight** » Wed Nov 27, 2019 8:28 pm

For some time sanrensei has been going out of favor, even before the advent of AI. And it seems like today's top bots don't like it, probably because they do not value the sides as much as humans have. In the Elf commentaries Elf may ding a side extension that used to be obvious by as much as 10-12%. Elf is extreme in its winrate estimates, but other bots agree that those moves are mistakes.

What about sanrensei? OC, the bots like nirensei, but that third stone on the side star point takes a hit. Elf usually rates it as a minor error (by my reckoning) losing in the range of 5-9%.

But I have just run across an exception, that I think is instructive. It is a game between Kosugi Tei, 4 dan, and Kitani Minoru, 5 dan, at the dawn of the New Fuseki in 1933, GoGoD 1933-10-17a. The first five moves maintain an almost constant winrate.

Click Here To Show Diagram Code: [go]$$Bc Kosugi Tei (W) vs. Kitani Minoru $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 2 . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . 1 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . 5 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . 3 . . . | $$ | . . 4 . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

As far as Elf is concerned, all of these moves are worth almost exactly the same, including

, which makes sanrensei. What is different about this board?

It's kind of obvious. It's the fact that

and

are 3-3 plays. Bots like to invade on the 3-3 versus the 4-4, 5-4, and 5-3 points. And they usually like to approach the 3-4 or use it to make an enclosure. But against the 3-3 there is nothing to invade, and I suppose that the enclosure or approach is not so big, either.

So my hypothesis is that it is not that Elf and other bots think that

is a worse play, per se, than any of these other plays, it is that against other corners Elf thinks that there are better plays than making sanrensei. Yes, the sanrensei was hardly on Elf's radar;

got only 30 playouts. Elf's top choice was a one space approach to one of the 3-3 stones, with 10.7k playouts. But the winrate estimates were approximately the same.

Perhaps it is not that the bots value the sides less than humans, but that they value the corners more.

ez4u · Post by **ez4u** » Wed Nov 27, 2019 9:17 pm

Don't forget that sanrensei was never actually in favor in any real sense. See the last chart in post #1 here. viewtopic.php?p=98494#p98494

Gomoto · Post by **Gomoto** » Wed Nov 27, 2019 10:04 pm

Who does not respect the half point advantage in fuseki is not worthy the one point advantage

Otherwise sanrensei is a sensible opening.

Bill Spight · Post by **Bill Spight** » Wed Nov 27, 2019 10:43 pm

Gomoto wrote:Who does not respect the half point advantage in fuseki is not worthy the one point advantage

A difference of ½%, even for Leela Zero, is almost surely noise. For Elf, it is even more surely noise.

Gomoto · Post by **Gomoto** » Wed Nov 27, 2019 10:48 pm

And for KataGo it is the world

I think there is a difference for pro level play and the trend the bots show is clear. But if you dont mind half a point advantage go on and have fun with sanrensei in amateur play. The problem with sanrensei is magnified by the mediocre follow up moves many players still choose. And these even matter in amateur level play from time to time.

John Fairbairn · Post by **John Fairbairn** » Thu Nov 28, 2019 3:32 am

Perhaps it is not that the bots value the sides less than humans, but that they value the corners more.

I would like to suggest an alternative hypothesis: the bots like probes more.

I have posted here before (well before AI days) about my surprise at discovering that what even we amateurs consider common-or-garden probes were rare and late among the Edo masters. For example, the under-attachment against the small knight shimari appears only once before 1800 and under 40 times before 1900. Similar remarks can be made about other probes, and I remember the astonishment in the whole go world when Yi Ch'ang-ho, again in pre-AI days, started making some weird new probes very early in the game.

Furthermore, and again as I have previously mentioned, the incidence of comments on probes (always in favour of) in professional commentaries seems very high. In the ~220 games in Genjo-Chitoku and Games of Shuei, there are around 150 mentions of probes. Note these are moves deemed worthy of comments - probes not commented on are not counted. It is interesting also to note that most mentions are in the later stages of a game, though great players such as Honinbo Shusai do occasionally stand out by recommending a probe even very early in the fuseki.

Perhaps the human dislike/antipathy/wariness of very early probes is that it increases complexity (or at least feels that way), and humans faced with an empty board crave simplicity. Bots don't have cravings.

I notice also that at least LZ faced with an empty board doesn't seem to look at all at 3-3 plays, but if you play a Black hoshi and then a White hoshi, for Black it suddenly brings the 3-3 invasion into the frame, even over a play in an empty corner (and there 3-3 plays are eschewed). Accepted this all seems to happen within a noise margin, but taken together with other considerations, it may be a genuine preference. Indeed, I think it is perfectly valid (and maybe better?) to consider other surprising AI moves such as early shoulder hits simply as early probes.

PS Just an hour after penning the above I started reading the latest Go World. Lo and behold, there is an article by Yokotsuka Riki 7-dan, in a series looking at AI tactics, entitled "The probe attachment". This refers to the following.

This tactic goes back to Jowa's day among humans, but only in the later stages of a game. Very early in the game it appears to be an AI-influenced play. I was amused to see this example: a probe of a probe (see lower left): DeepZen 's move against Shibano Toramaru (admiitedly not really an "early" example, but still may something about the bot "mindset".

I was amused

Gomoto · Post by **Gomoto** » Thu Nov 28, 2019 6:05 am

plain 3-3 without hoshi is lacking efficiency

gowan · Post by **gowan** » Thu Nov 28, 2019 9:56 am

I think many amateurs play by recognizing patterns. When play deviates from familiar patterns it immediately becomes more difficult. Maybe we've all experienced having our partner in a game make a move that we feel sure is not good but we don't know how to handle it. I think that was probably the case at the beginning of the shinfuseki period and now at the beginning of the AI-influenced period. Now it seems that sanrensei is in decline. I remember a time when learning how to play white against the sanrensei opening was a necessary point of growth for SDK's on internet servers. Now everyone, including pros too, has to come to terms with AI-type moves. Just as with the shinfuseki moves 90 years ago now it will take some time before the AI approach to playing is better understood and integrated into the general go community. Meanwhile, amateurs who understand basic go principles will be able to take advantage of opponents who play AI-type moves without really understanding them.

TelegraphGo · Post by **TelegraphGo** » Thu Nov 28, 2019 10:40 am

Gomoto wrote:plain 3-3 without hoshi is lacking efficiency

Strongly disagree. If the san-san opening was a bad opening, it would be because the followups are inefficient, not the first move. The first move is the most efficient way to take the corner ever! Also, it isn't a bad opening, usually losing only tiny percentages to public AI evaluation, and often the best first move in the corner when there's some opposing stones in the sides surrounding it. Moves that some versions of AlphaGo play are pretty much never going to be "lacking efficiency".

Bill Spight · Post by **Bill Spight** » Thu Nov 28, 2019 11:15 am

TelegraphGo wrote:
Gomoto wrote:plain 3-3 without hoshi is lacking efficiency
Strongly disagree. If the san-san opening was a bad opening, it would be because the followups are inefficient, not the first move. The first move is the most efficient way to take the corner ever! Also, it isn't a bad opening, usually losing only tiny percentages to public AI evaluation, and often the best first move in the corner when there's some opposing stones in the sides surrounding it. Moves that some versions of AlphaGo play are pretty much never going to be "lacking efficiency".

Elf says that the 3-3 on Black's first play loses about 3% versus the 4-4. But as

in the adjacent corner to Black's 4-4, it loses only ½%. For

even if it is not a mistake, Elf has a clear preference against it. Not so for

. AlphaGo Teach, which has a smaller acceptable range than Elf, thinks that the 3-3 loses 0.6-0.7% in either case. Definitely playable.

Gomoto · Post by **Gomoto** » Thu Nov 28, 2019 2:59 pm

@TelegraphGo

I was comparing these two moves:

plain 3-3

Click Here To Show Diagram Code: [go]$$B $$ +-------------------- $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . 1 . . . . . . . $$ | . . . . . . . . . , $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . , . . . . . ,[/go]

3-3 under hoshi

Click Here To Show Diagram Code: [go]$$W $$ +-------------------- $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . 2 . . . . . . . $$ | . . . O . . . . . , $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . . . . . . . . $$ | . . . , . . . . . ,[/go]

Move 1 is less efficient than Move 2 (Numbers only used to identify the two plays). This was my statement and I stand my ground

.

Both are sensible opening moves.

TelegraphGo · Post by **TelegraphGo** » Thu Nov 28, 2019 4:54 pm

Move 1 is about as stable a situation as there is, and Move 2 about as unstable as possible. They're hard to compare. I still don't see any reasoning myself that move 2 is more efficient. If you mean that you immediately take more territory with it, you're right - but you also give up considerably more influence. So the overall balance may well be fine. As far as I can see, it's dependent on the global situation (mostly how much can white use the influence from Move 2) to tell which is more efficient.

I would generally still expect move 1 to be a little more efficient in general, because otherwise if 3-3 is close to as good as 4-4 and 3-4 (almost always true), then you should invade 3-3s before taking empty corners. Some AIs are doing that, but as I understand it, mostly as a probe. If you had to compare who was better if the whole board was black taking sente variations of getting 3-3 invaded by white in each corner, I'd guess that the side with influence is better, even to AI. Maybe I'm wrong, and need to reevaluate.

As something of an on-topic aside.
My understanding of 3-3 by itself is that it's a strong move, but sometimes gives your opponent more control over the game, since you'll have a hard time doing anything efficient nearby. That's core to my reasoning for why sanrensei is fine against dual 3-3. If it was just corners before sides, like Bill suggested, I would expect a knight's move or large knight's move enclosure to still be 3-4% better than the middle move. So I think this is a really early example of AI valuing 'initiative' in games - by setting up sanrensei [explicitly when W can't happily build anything of their own], B is likely to get more control of the game (loosely defined as W needing to answer B more than B needs to answer W) later on. AI's seem to put a little extra value on positions with this kind of control, more than humans have, so I think maybe AI's judge the side move as a side-over-corner inefficiency invested into getting extra control later. An even tradeoff, or so ELF seems to claim.

Gomoto · Post by **Gomoto** » Thu Nov 28, 2019 5:43 pm

I think of move 2 as the move that is achieving (in a subtle way) more things on the board at once. It feels more like a kind of a multipurpose move to me. (What this purpose will be is not decided yet, so you may call this a probe as well if you like.)

Bill Spight · Post by **Bill Spight** » Fri Nov 29, 2019 2:07 am

TelegraphGo wrote:As something of an on-topic aside.
My understanding of 3-3 by itself is that it's a strong move, but sometimes gives your opponent more control over the game, since you'll have a hard time doing anything efficient nearby. That's core to my reasoning for why sanrensei is fine against dual 3-3. If it was just corners before sides, like Bill suggested, I would expect a knight's move or large knight's move enclosure to still be 3-4% better than the middle move.

I checked with AlphaGo Teach and it does prefer the keima enclosure, the one approaching the White side, over the sanrensei by 1%, 47.1% to 46.1%. I expect that that is within its margin of error, but I think it is a definite preference.

Kirby · Post by **Kirby** » Sat Nov 30, 2019 2:59 pm

In response to the OP, perhaps it’d be feasible to play double 3-3 opening and then revert to how we played before AI

Life In 19x19

So what's wrong with sanrensei?

So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?

Re: So what's wrong with sanrensei?