AI based score estimation and auto-scoring is now live

[stone.defender]

@yebellz

one could use the analysis feature to play a move and then ask the strong AI to estimate the score for that new position.

@yebellz

Further, if one makes enough queries to the estimator (which is even easier to do with correspondence timing), then one could even figure out what the AI thinks is the best move.

Whoops: @anoek, you just implemented way to ask bot where to move next.
choice of both: AI score estimator and analysis shouldn’t be allowed in ranked.
Only one of the two.

better to create very easy to use graphical editor that counts geometrical area

@stone.defender

I would prefer score estimate where you have huge brush tool like in Paint graphical editor. So it counts geometrical area of what you painted.
So no AI tells you additional Go information, but you can count very easily in geometrical sense.

[anoek]

Area scoring

[Feijoa]

It looks like handicap counting is still not correct for Chinese rules. Example game in which White has just won by 13.5:

[stone.defender]

@stone.defender

Whoops: @anoek, you just implemented way to ask bot where to move next.
choice of both: AI score estimator and analysis shouldn’t be allowed in ranked.
Only one of the two.

@anoek, there is easy solution actually:
disable estimate in position that you created in analysis, AI estimate should work only in actual game position

[_KoBa]

Yeah, estimator is incorrectly giving black points for handicap stones - that “bug” has always been there. Its not anything new, ppl just have to realise that estimator gets handicaps and sekis wrong >___>

[Popz]

this is such a bad idea…
especially if disabling it doesn’t disable it for spectators. All people need to do is to open the game in a private window to start checking what kataGo thinks of the groups l&d status…

please, if you still go ahead with this, at least allow the game owner to completely disable it for everyone.

[Groin]

that’s a great idea, using the specificity of an online game (vs real life) and not going too far on the external resources side

[stone.defender]

I was wondering why estimate is so different from Kata review

But it seems, its just woks completely incorrect

[DVbS78rkR7NVe]

And it even correctly tells me to live if I play away.

And big squares at H8 and H1 hint me what to play. The miracle of technology.

[shinuito]

@stone.defender

But it seems, its just woks completely incorrect

@anoek

mm yeah that’s an artifact of me hard counting area that’s not really territory yet - probably better to use the point estimate that KataGo gives back now, I’ll look at updating that.

^^ for finished games at least

[wordmau3]

Yes, on mostly empty boards, something seems to go wrong. For the empty board, I get scores between b+250 and b+350

Edit: Looking at the beginning of the thread, this observation is likely nothing new

[anoek]

PSA: I’ve disabled the new SE for players of ongoing games and instead use the old SE, so effectively no change for ongoing games. For uploaded games, reviews, or spectated games, this new system will be used.

This is an interim fix while I work on figuring out exactly when the SE should be usable and not, and add optionality etc.

[DVbS78rkR7NVe]

Yeah, I think you can try to figure out best move from the eval of the current position.

Look, here, black to move, it suggests that black stones aren’t completely dead and black should try with E10 cut. And before you say it’s obvious, my 2k opponent didn’t do it.

[_KoBa]

@S_Alexander

Anyone got complicated triple-ko seki positions?

it looks like it actually can NOT handle triple-ko’s and other really complicated things where everything relies on whose move it is. Couple of examples where SE seems to get things wrong :

triple-ko Friendly Match

triple-ko il2019 vs. int darth_vejder

[yebellz]

@stone.defender

I would prefer score estimate where you have huge brush tool like in Paint graphical editor. So it counts geometrical area of what you painted.
So no AI tells you additional Go information, but you can count very easily in geometrical sense.

I think this could be better accomplished by allowing one to mark dead stones in analysis mode.

The analysis already does count territory and give you to the score, except that all territory borders have to be closed and all dead stones need to first be captured (under Japanese rules, you’d have to avoid passing with both sides to prevent distorting the score).

Basically, my proposal is that instead of painting all of your territory to count, you should instead just be able to finish the borders in an analysis branch and mark the dead stones.

[stone.defender]

@yebellz

this could be better accomplished by allowing one to mark dead stones in analysis mode

yes, but it worked weird

@yebellz

instead of painting all of your territory to count, you should instead just be able to finish the borders

you could do that but it was very slow to click each time to place stone. It should be possible to seal borders as easy as in graphical redactor

[yebellz]

Yeah, making the specific interface smooth and feel friendly is a bit tricky.

One could similarly “paint” the border lines closed, and then clicking (or painting) any stones in a closed off area should automatically mark all of those as dead.

However, what I don’t want is to have to paint (like a brush) over ALL of my territory just to count. Maybe it should be more like a “bucket fill”.

[Vsotvep]

@pdg137

Here’s a specific example of how the new score estimator could give away a game:

It seems like this kind of thing is not going to be solved by disabling the estimator until the end of the game, since presumably if we had passed at this point, auto-scoring would have revealed the weaknesses.

Let’s see how we could resolve this problem, since as a moderator having to fix a lot of wrongly scored games I would really love to have a better auto-score at the end of the game without giving away the position to the players (which is only going to give us more reports about “opponent resumed the game after learning what to play from the scoring stage”).

---

In the situation above, we assume the two players both passed and consider the one White stone to be dead. KataGo, however, considers the Black stones to be dead if White is to move, and the White stone to be dead if Black is to move. This territory should therefore be easy to spot automatically: if certain intersections swap sides depending on whether Black or White has the next move, they are unsettled.

Now, to whom should KataGo grant the unsettled area? I think we could solve the majority of these problems by making two distinctions:

• There’s the “open border”, where it’s possible from an unsettled empty intersection to reach settled stones of both players. Such areas should be marked as dame, and the game should in theory not be scored and be resumed before concluding.
• There’s the “killable group”, where a group needs an extra move to avoid being killed. In such a case, the stones of the “dead” group usually outnumber the stones of the “killer” group. This is the kind of position we have at the top: 4 black stones are threatened by 1 white stone. In such a case, I suggest that KataGo flags such an area as belonging to the colour with the most stones inside the unsettled area.

[_KoBa]

Wait its doing fractions now?