Question about OGS Joseki position (let's keep meta discussion in the other thread)

I noticed that in the well-known https://online-go.com/joseki/63, B is marked as only good and not ideal. Is there support for this view? I did get the impression that its popularity is in decline, but in sufficient extent to warrant demotion? (Personally, I don’t consider it ideal in my kyu opinion, but it’s not up to me to make that call.)

Unlike the other thread, I don’t have any preconceived opinion about this; I genuinely don’t know whether it should be called good or ideal.

AlphaGo Master pointed out first that B is suboptimal (because of the aji of a P15 peep later) and pros convinced themselves that AlphaGo is right about this. There are several pro videos about it, like the one linked in the comment and earlier ones where Michael Redmond 9p comments on AlphaGo’s games.

A is played 3 times as often as B in Waltheri’s (search window including starpoints on both sides) and most of the games with B are from 2017, when pros were still learning about the new AI variations from an early 3-3 invasion.

I think that B being not ideal is even the reason why these early 3-3 invasions are considered playable (ideal even) nowadays.

So I feel that B is marked correctly here.

You might even say that perhaps black’s hane at P17 is already slightly suboptimal, considering that in position https://online-go.com/joseki/65, only AlphaGo Lee plays C. AlphaGo Master usually plays A and AlphaGo Zero usually plays B. But there is a question if top pros are actually strong enough for this fine level of distinction to matter.

3 Likes

It had been discussed a couple of times.

The reason it ended up being good is because pros choose it less than 3% of the time. It is one of the more borderline ones especially when you’re asking AI and it’s losing like 1% to play. (for perspective, some AI judge a lot of joseki responses to approach moves to lose up to 5%)

There also have been a number of professionals talking about why the new joseki are a bit better.

edit: besides it being played just 3% after already deciding to hane. You also have to compare that joseki to black’s other responses to a 3-3 invasion. If you do that, then professionals choose that joseki 0.2% of the time.

3 Likes

I wonder if score loss with Katago is better than win percentage.

The win percentage drop I imagine to be more sensitive than score loss. Losing 0.5 of a point in a corner wouldn’t seem to bad but it could be a big loss in percentage in a close game no?

I don’t run Katago on a pc myself, just use OGS or aisensei to review my games. Would you have to test lots of different combinations of 3-4 and star points in corners when testing out a joseki and looking at win percentage or?

I think loading up a bunch of pro games where the joseki was played and then analyzing it with a couple of different weights/programs is the best way to investigate that. I fear that people just use blank board or analyze it for the first joseki of the game…

Actually it is really obvious that some AI patterns on josekipedia are done on blank boards (and done by Leela from two years ago because they don’t know ladders :face_with_raised_eyebrow:)

Well… Maybe doing both of the above is the best? You can’t separate the josekis properly from the games if you only analyse games. Typically the choice of second/third joseki etc is depending on what’s happening elsewhere in the games. For instance even watching Michael Redmonds videos about 3-3 invasions, the choice of which one to use can depend on the rest of the board. So you can’t really say conclusively that move ‘x’ is a 2% drop in winrate if tends to be played later in the game or in certain positions?

I don’t think ladders are that problematic though. I think once you tell people this is good if you have the ladder or only playable with a ladder etc it’s fine no? It does bring up the question like https://online-go.com/joseki/11117 ‘D is good when this ladder is favourable’. I can’t remember the result of the discussion of how moves should be categorized as good/ideal if they depend on a ladder?

Thanks guys, I guess it’s considered subop’ and in the right category.