Mark5000's Guide to the Opening (v2)


#1

Hello OGS. This is an update to the prior thread mark5000's Guide to the Opening. This lesson is intended for novice players who would like to understand high-level opening theory presented in an accessible way. This theory includes innovations of AlphaGo, which rendered the previous guide somewhat outdated. As before, I don’t claim to have any special knowledge about the opening. This is just a distillation of recent Go knowledge through the eyes of a Go lover and enthusiast. I hope it helps you understand the opening better.

OGS demo format

SGF for quick access and offline viewing

mark5000’s Guide v2.sgf (19.8 KB)

Please let me know if there are any errors or necessary omissions or any other thoughts you may have.


Using leela zero to estimate win probability of a board
#2

Thank you for updating your wonderful guide!

The impact of strong AI on opening theory sounds very interesting to me, however, perhaps I’m not strong enough to really appreciate what we may have learned from these bots. In what ways has AlphaGo most impacted our understanding of opening theory? Could you point out some specific examples where it has really changed the way we understand certain positions?

Sorry for the redundant questions if you already go into this in detail in your guide. I haven’t had a chance to work through it yet, but I am really looking forward it!


#4

Sure. In my view, here’s what the key innovations are, and what’s changed from v1:

  • The axiom corners-sides-center is now just corners. Corners are more valuable than once thought, and splitting moves on the side have vanished. This means that opening play centers around the corners, after which strength and weakness of groups determines play more so than location on the board.

  • Named openings are dead. We all know there are a huge number of perfectly valid moves in the opening. But there was an effort, treated more religiously by westerners, to label certain openings similar to how chess openings are labeled, such as orthodox, Chinese, Kobayashi, etc. We now know these named openings are not special, and the AI-suggested “best” paths avoid all of them entirely. I preserved some of the lingo in the lesson, but mostly it’s back to the drawing board.

  • The 3-3 invasion reigns supreme. A consequence of increased value of corners is that the 3-3 point invasion of a 4-4 point is now a first-order move, just behind taking a corner in terms of priority. This means it can be played as early as move 3. The joseki have also changed, including the beginner-friendly bend-and-extend joseki, which now won’t give the opponent a hanging connection.

  • Answering a 4-4 point approach is just big not urgent. There’s new ways to handle the double approach of a 4-4 stone, namely by attaching on top of either stone. Since this gives a better result than once thought, approaching a 4-4 point can be gote, which is another reason why the 3-3 invasion has increased in popularity.

  • Enclosure popularity is reversed. Pro and amateur players used to choose the small knight enclosure of a 3-4 point more than any other enclosure. To everyone’s surprise, the two-point high enclosure is the preferred enclosure of all bots from AlphaGo to Fine Art to LeelaZero, and the small knight enclosure is the least favored one, being slightly overconcentrated and vulnerable to shoulder hits and attachments.


#5

It’d be nice to have numbers marking all of the continuations for the OGS version. In the empty-board starting position, for example, marking the 3-3, 3-4, 4-4, 3-5, and 5-4 would be a nice visual indicator of what the different branches are. There’s a few instances where some, but not all, of the lines are labeled with letters in OGS, such as after white plays the far corner, where black’s near corner responses are marked but the diagonal is not.

The eidogo SGF sort of does this, but transpositions aren’t marked. For example, black and white play hoshi in opposite corners, and black follows up with nirensei, there’s no marked continuation. Playing the last hoshi transposes back to a commented line, but having a visual indication of that would be nice.

Also, not sure why, but I get a different variation if I open up your review vs. if I download the SGF. The downloaded one seems like it has more variations, and is likely the “v2” version of the guide.


#6

I hear you. It’s mostly a shortcoming of OGS. When you download a demo SGF from OGS, that SGF contains all branches ever created in the demo, even if you deleted them. It also sticks with the original branch ordering, even if you reordered them in OGS. I had to download the SGF and then manually prune the branches I deleted in OGS, but it still doesn’t match perfectly. I’m not sure what to do about that.


#7

Thank you making this list. I have often heard people remark generally and vaguely about how the bots are changing our understanding of opening theory, but I have never seen such a clear and descriptive list of prominent examples. I was unaware of the scale and nature of the impact, and your list is very helpful to me.


#8

I have always, instinctively been especially keen on taking and keeping corners. I feel validated by this information. Is it accurate to say that the bots also emphasize a territorial game rather than an influence/moyo game?


#9

@mark5000 Great summary of the recent evolutions in opening theory!


#10

Something I struggle with is the question “does this knowledge apply to me (a solid DDK)?”.

For example, when the 3-3 invasion thing first started hitting, Dwyrin said basically “that’s nice if you can do whole board reductions skillfully, otherwise maybe not, for us mere mortals”.

Is that still the thinking? Or is the knowledge of what to do with these innovations getting well enough understood to pass down the ranks?

I suspect there’s something similar with two-point-high 3-4 enclosure preference. That might be important if your opponent is accomplished at attachment play … but are they? Is the bot-perceived drawback of low knight not actually applicable at DDK level of play?

At the meta level, this is a topic I’ve been interested in for some time: the question of how real and significant “stratification of strategy among ranks” is. It became more interesting with AIs, because they far extended the spectrum of strata…

I don’t see much in the way of formalisation of this in Go teaching and thinking - the tendency is more towards expecting that “a good move is a good move”. I’m not at all sure that’s true, and I suspect that at some point our approach to training may stratify more…

GaJ


#11

I don’t think so. AI has stunned us with both bold invasions and bold fifth line plays.

As I have mentioned in another thread, the simple fact that they favor the 4-4 point in the opening and allow for a 3-3 invasion shows that they are not too territory oriented.

Just to be clear, @mark5000 never talked about keeping corners. “Corners are more valuable” should not be interpreted as taking corner territory.


#12

Cautiously, I would say yes, zero-based AI are less biased towards large frameworks than many humans and certainly the former generations of AI. The caution is because strategy and tactics are closely intertwined. Staking out a large territory isn’t viable if your opponent can destroy it, and zero-based AI is exceptional at destroying influence. So AI does not necessarily justify the territorial playing habits of a human kyu player.

This also relates to @GreenAsJade’s thinking. Since humans can’t play with the same unwavering tactical precision as AI, does it make sense for a person to mimic AI strategies, given that the changes in strategy were the product of sharper tactics the person does not have? As any good lawyer would tell you, it depends. For example, if your opponent, as Black, plays the orthodox opening, it may, for you, be a mistake to copy AI and let it happen, if you can’t deal with the formation effectively. On the other hand, you could let it happen, and then learn from the experience. And that is where I feel strong AI comes in: it encourages players of all levels to stretch their limits even further and play moves that were once out of their imaginations, because more moves are possible than we once believed.


#13

That’s a great observation!

I’m still trying to get traction for my instinct that the way we talk about good moves and especially teaching needs to be more stratified and skill-contextual. But the point you made does illustrate why I think this needs to happen.

So for example, there must be a “strata” where the insight “it is no longer corners-sides-centre just corners” applies, but a “lower one” where surely “corners-sides-centre” is still the sound basis for play.


#14

not only dwyrin is of that opinion. even michael redmond has stepped back from trying to emulate AI play somewhat, because he feels unable to do it well enough.

well, that is true i think. its very pssible to provide solid advice to a beginner even when we know that it wont lead to perfect play. it is meant to help a student of the game to play better, not perfect. therefore the advice mainly has to be helpful, not perfect :smiley:. sometimes a sub par answer that we can reason for or against might even be more helpful than just providing the solution. there are definitely things worth teaching a beginner, even after the AI have changed our view on the game.
that depends on a number of things though. for one, it depends on the outlook of the student on the game. if teaching a prospect pro, i would try to teach what we (and the AI) think is best, not easiest. for a western player encountering the game in their mid 20s this might reverse. the fastest progress is likely to occurr by teaching what is easily understood in terms of larger concepts even if that leads to incorrect conclusions sometimes or encourages a tradeoff of perfection in favour of avoiding complications.
it is also obviously much easier for teachers to teach what they think they understand. since we are in a phase of huge change there arent (m)any teachers comfortable teaching the new style yet, especially since their own results using the style were lackluster. in my opinion though, this might still change as well.


thank you very much for this great compilation and analysis @mark5000!


#15

Mark5000, I just went through your tutorial and found it highly informative and useful. Thanks so much for your effort on this.

I do have a question on one aspect already. Your recommend opening strategies concentrating on the corners, and your examples are on 19X19 boards. But I have just read in two different places about openings on a 9X9 board. They both recommend opening on tengen (the center of the board). They cite different strategies of play on a 9X9, even among some advanced players, such as 6kyu or even dan players.

I’d love to hear your reaction to this. Does the strategy on a 9X9 board differ so much from a 19X19 as to necessitate a different opening strategy? Or indeed, is this that much different, since tengen on a 9X9 board is about the same distance from the corner as hoshi on a 19X19.

Any comments?


#16

9X9 opening strategy is completely different. The entire game is about fighting, so playing at or near tengen is optimal because it lets you fight more easily. Developing corner territory isn’t really a thing. Joseki aren’t really a thing.


#17

@mark5000 has already done a very in depth discussion about 9x9 strategy in another thread


#18

Yes, 9x9 openings differ because the entire board consists of overlapping corners. For 9x9 move 1, the three center moves are optimal because they occupy or influence all four corners at once. A central position is also important for the close combat that often follows in 9x9 play.


#19

Thanks. What a supportive community!


#20

Sorry, but I don’t want to play like pointless piece of iron. :sunglasses:


#21

Not optimal enough for you? :wink: