I think the “fully covered” option is much better as it follows the same idea of territory. Here is W’s territory, here is B’s territory, and here is a disputed territory that you need to pay attention to.
I don’t think it’s neccessary to have a distinct “seki” status to mark groups during scoring. Seki is just alive (albeit without territory under Japanese scoring rules, but that is taken care of automatically by the upcoming seki detector). I think requiring players to mark sekis would make group status marking unnecesarily complicated. And for what purpose?
Just the “dead” marking - where “not dead” means alive (possibly in seki) - seems sufficient even in this complicated case:
And also in other potentially more difficult cases, like below, it seems sufficient to only mark which stones are dead (and the upcoming autoscore marks those correctly):
Are you sure though, that this is what actually happened? The AI applied full fledged human psychology to the state of the Goban to come to the conclusion that both players most likely thought the groups were dead and so the AI marked them as dead?
Can someone who was involved in the development shed some light on this? Is this an error or does the AI actually take the player’s probable thoughts and feelings into consideration?
1 Like
Using a Go Magic puzzle that a confused beginner posted on Reddit as a test case on the beta server (presumably Japanese rules):
Upcoming autoscore marked correctly (as I’ve grown to expect by now):
(B+4.5)
But the beginner was thinking that black’s A19 stones and white’s T12 stone should be marked dead also, because they can be captured.
I can get their confusion, but their assessment happens to not coincide with the regular marking of dead groups. Nevertheless, it wouldn’t matter for the net score, because removing those living stones would increase the territory points of the victim by the same amount as it would increase the prisoner points of the adversary. So it’s wrong, but inconsequential.
When I tried this on the beta server by marking those capturable (yet living) stones dead, indeed it did not affect the net score. So that’s fine:
(B+4.5)
I was also testing this because I was curious if the upcoming marking process would allow 2 opponent dead groups “touching”. From the above screenshot it seems that it does.
Even though it is inconsequential for the net score, should that even be possible by the rules of go? Or should the marking process automatically unmark a dead group when some “touching” opponent group is marked dead?
I feel the latter would be proper behaviour, but it would complicate matters when 2 opposing dead groups “touch” via an area that is neutral by open borders, like this:
I’m pretty sure it doesn’t model the thoughts of human beginners. In this case, the autoscore just happens to agree with my educated guess (as a human player and teacher) about these beginners’ mistaken assessments.
But as in many cases where the game stopped while some group was still unsettled, we can’t be sure what both players were thinking and if they even agree.
Also in this case I’m not sure. Was it an actual beginner game that stopped at this point, or was it just a test position that you set up?
To be clear, I prefer the thin line to just a sea of red, I don’t think a sea of red is going to do any favors to those that actually hit the sealing system (which is beginners, I don’t expect anyone commenting in this thread to see it in their own games).
Now, if I could code it reliably, I would prefer something like the marking of n6 as you have, but I tried a few things that came to mind and I couldn’t come up with a reliable general purpose algorithm that did smart things in all cases. So I think the marking of adjacent stones is good enough, and superior to a sea of red for the purpose it’s trying to fill. Still, if someone comes along with a better algorithm for highlighting areas that need to be sealed, that’d be swell.
6 Likes
I also prefer the thin red line, both practically and aesthetically
1 Like
Yes, that’s what I think as well. But then I strongly disagree with your assessment of the situation:
If the erroneous ouput randomly aligns with reality then the AI did not actually do a good job here. It infact did a bad job because it should not have marked those groups as dead.
All the games I tested were actual beginner games from when I first started playing Baduk three years ago. I was White in that game but I can not remember my thought process from three years ago.
I agree with you on this one. The sea of red does seem to be more logical and correct to a Baduk veteran. But it might confuse the players that actually need it.
EDIT: There is also the “slightly fuzzy” option though. That one seems the best to me because it clearly depicts the area of action (which makes it better than the unhelpful sea of read) while it doesn’t invite silly moves (like the current implementation that tells the player to play L1).
2 Likes
I haven’t actually tested this, but I’d assume that applying some image processing kernel like edge detection on the ownership field of a contested area might help to come up with good locations for open border markings.
Yeah, the intent here is that the players thought it was alive so should be left alive. KataGo knowingly belives the group is dead if white players first, alive if black plays first, so our logic that uses katago just leaves it alone, it’s a later phase in our stone cleanup process that removes it, I’ll get it fixed up. Thanks for the catch.
3 Likes
Perhaps so, but it’s definitely a bit of a research project and not guaranteed to have a favorable outcome. I welcome contributions for someone that wants to explore that more though.
4 Likes
I tested what happens when that black group is marked alive, hoping to see open border marking (perhaps at A7), but no:
(also not when I mark B7 dead, nor when I add a black move at A7 before passing)
This game is quite close to being properly finished and the gap in white’s border is quite small, so perhaps it is unexpected that the current open border implementation does not mark it?
1 Like
Excellent!
Also when black connects at A7 and the game stops, the group status and open border are now marked properly:
And when white then closes the border before the game stops, black’s unsettled group remains alive as it should:
Thumbs up!
4 Likes
To be sure, I tested some more positions where play in that corner continued a bit further, and the latest autoscore seems to handle all of them as intended:
Edit The marking of this unsettled position surprised me a bit.
Black can make a seki by playing at B3, so I expected the autoscore to mark both black and white alive in this corner.
Is this perhaps too much of an edge case?
3 Likes
Hmm, yeah, this seems to be a good edge case test for the algorithm to clear unsettled stones - I’ll work on it more in the morning
2 Likes
More experimenting with this corner shape:
Unsettled because black can live in seki with either atari, while white can kill with A1. Autoscore marks black alive (which is good), but perhaps it should also mark white B2 alive, because black can’t kill it even if black plays first?.
When white plays at A1 before stopping, I don’t quite understand the marking below. It seems that autoscore thinks black is still unsettled, but isn’t black unconditionally dead here?
I think black is unconditionally dead because black can’t do much in hypothetical play, while white can prove black dead like this (and then white A2):
Though perhaps autoscore is aware that black is unconditionally dead there, because when stopping before black B3, it apparently doesn’t see a way for black to survive:
That autoscore marking seems correct to me.
Or am I overlooking something here?
3 Likes
I think you are right, but I want to highlight that this is the sort of shape that some might find contentious, as discussed earlier, about whether the scoring system is providing hints to the players.
It is possible that both players have passed at this position with the assumption that Black (in the lower left) is dead, or they might have both believed that the two White stones are dead, or even yet they might be in disagreement about the status. It comes back to the debate about how much should the scoring phase should provide the convenience of marking stones versus the concern that it might provide undue assistance to the players.
Under area scoring rules, in general, if both players agree on the status, even if strategically wrong, then that should be kept as the case.
3 Likes
I think that’s also the intention under territory scoring rules. But I also think that we’re aiming for the upcoming autoscore to only mark some unsettled groups as alive, not dead groups. AFAIU, in properly finished games where nothing has unsettled status, autoscore should still mark fully alive stones as alive and fully dead stones as dead.
But this position looks like a properly finished game to me (unless I’m missing something), where black’s corner group should be marked dead and white’s corner stones should be marked alive (under Chinese rules as well as Japanese rules). But autoscore did the opposite.
3 Likes
I imagine we’d allow such markings even when the unsettled-ness is for normal rules,
rather than limiting it to when the unsettled-ness also applies to hypothetical play.