Does the AI winrate simply tell us what is "Joseki"

Just to add my resounding “yes” to the other two above.

This is the current “policy” of OGS Joseki Explorer: Joseki need credible (typically professional) sources.

The concrete reason for this is that AI does not make a judgement about “Joseki” at all. It doesn’t even know what one is.

The definition of “what is Joseki” is a human construct, and humans need to assess whether a sequence meets the criteria. We hav decided to require this to be a Pro, or equivalent.

During discussions about this topic in the Joseki Contributors group, it’s emerged that exactly what that criteria is is kind of loose - you know it when you see it, but it defies precise definition.

Nonetheless - the assessment that a sequence of corner moves can be considered “generally optimal all other things being equal” has always been something we’ve relied on professionals for, and still is.

(Among many things it doesn’t know about, the AI doesn’t know what “all other things being equal” means).

Footnote: it’s not that we don’t “trust the ideas of the AI” as you put it. That’s a very general statement. Specifically what we don’t believe is that just because the AI says a move in a sequence has the highest winrate, means it’s joseki, and just because the AI plays a sequence more than others, that also doesn’t mean on it’s own that this sequence is “Joseki”.

5 Likes

But if an AI is playing itself and has played millions of games to conclude that a sequence of (opening) moves is equal and optimal, this is, IMHO, joseki by definition. What is not clear is that players who are not AI would benefit from seeing these sequences.

But the trick is that for the AI, the sequence that is optimal is optimal for that board position.

This is not what “Joseki” is.

Joseki is “optimal, all other things being equal”. Which as I said is a loose phrase that captures the sense that we feel that this sequence can be relied upon across a wide range of board positions, as long is there isn’t some prevailing factor.

I do agree that if we see the AI playing a particular sequence again and again, in varying circumstances, then it is a candidate for Joseki, and indeed this is what the pros do: they look at AI games and adopt sequences that work in a range of situations. That is what leads to the current revolution in Joseki.

But this is different to mechanically running lots of AI and picking common sequences from those games and blindly putting them in a “dictionary”. This is the kind of suggestion that led to this thread.

That would be one kind of tool, but it is not the kind we are building.

In contrast, we are building a dictionary that captures very useful human pro filtering of what the AI is showing us, such as this kind of great stuff and presenting it to us amateurs in hopefully a useful way.

2 Likes

I think the purpose of a corner pattern dictionary is to play better Go. I play a game and get a bad result, where did I go wrong? Or I’m playing in a tournament soon I need to review some common sequences I might actually encounter in my games.

IMO people get stuck trying to define the word Joseki. Use the mission statement above to guide what variations to put, and what commentary to add.

1 Like

I agree with the purpose.

The “purpose” of the Joseki Dictionary is to help us play better go, for sure.

In a specific way: by providing us high quality content in an accessible way.

But we aren’t defining the word “Joseki” for semantics.

We are doing that because defining the word joseki is a shorthand for “defining what meets the mission statement”.

I guess there is some value in asking the question “does the useful content we include have to meet the definition of Joseki?”, and this is what you are saying.

At the moment it seems like these two things overlap so strongly, and “is it Joseki?” is a useful yardstick.

Note that it’s not the only criteria - we include “Tricks” for example, and mistakes and refutations.

I think if we were to say “the heck with it, forget about Joseki and the meaning of that, just ask does it meet the mission statement”, we’d be having the same discussion minus the label.

1 Like