if only there were an option “failed to autoscore, you’re on your own folks”
2 Likes
That could be even smarter. “Please resume playing” or “can’t score: game annulation” choice when manual scoring is not available.
Another way to look at it is, what is the bigger hint?
- Revealing the seki, when they saw the atari
- Revealing the atari, when they saw the seki
Telling then to resume is guaranteed to give them a hint no matter what they thought, so that doesn’t seem like a great solution.
2 Likes
how does one codify that? Like sure it’s trivial to look for atari, but do we have to look for heuristics like atari/potential seki and then weigh them based on how easy they are to find?
1 Like
Autoscore v3 does it by removing unsettled stones entirely surrounded by the other color.
Sure, “atari = dead” could be another heuristic. Or you could sort through all the 2n choices for the unsettled groups and pick one that minimizes seki.
But even if there’s no great algorithm it’s still good to know about the failure cases of this one.
1 Like
So in this case, v3 will remove the stones (even without atari rule)?
I think that’s reasonable - at least it doesn’t leak information
1 Like
BTW I remember seeing recently an autoscore fail with this shape.
It leaks the information that white is not unconditionally alive.
1 Like
okay fine, but practically, I find it really hard to believe two players think it’s alive. (if they did, then white would extend anyway, since borders are not closed!)
Edit: So i think the 3 cases are:
- undecided → white should extend or black should capture
- white is unconditionally alive in seki → white should extend to close borders
- black can unconditionally capture → black should not capture, since he will lose a point
IMO the last is the only logically consistent one if both players have agreed to end. Or maybe I’ve missed the one where “white can unconditionally capture → shouldn’t capture to avoid losing a point”. But again, pragmatically this feels unlikely.
1 Like
All sekis I know require that you don’t close all shared liberties.
2 Likes
This is kind of off-topic but maybe this is a seki with open borders?
game link | autoscore tester
Autoscore seems to handle it OK, doesn’t even warn about the open border.
2 Likes
Is this anything?
I think it’s unfair to expect the autoscorer to intelligently handle it, but… kind of felt like this:
1 Like
Maybe, do you have a game link?
That’s weird that the upper left black territory is not marked.
asdf vs. asdfg I played past it, but I think it’s reproducible (at least I hit “Auto-score” multiple times)
Seems like the estimator understands that if white makes the right move, the top left black group is also dead (but I guess you mentioned some special handling of totally surrounded space, so that doesn’t make sense)
is neural net used in scoring of so small boards?
1 Like
I guess it just marks everything alive which seems OK if it’s unsettled:
Interestingly my v3 algorithm chooses to mark everything dead in this case.
For this thread I wasn’t really thinking about how the territories get marked since I have no idea where that’s coming from and I think it’s a second level on top of autoscore. The autoscore API endpoint just returns the live/dead status of the stones (and rarely, unsettled).
We don’t have any hard info but I would be really surprised if Anoek coded a completely different algorithm for 5x5.
1 Like
Its only used for 9x9, 13x13, and 19x19 afaik
With unstandard board sizes, players still need to manually click all the stones alive/dead.
this is clearly wrong if you look at the discussion
Ah so it seems!
Yeah, tho sometimes it fails big time on unstandard boards. I just finished this 25x25 game, where everything was transparent at first, had to mark everything manually. And i was unable to mark the top left territory and the A17 B16 D8 correctly >___>
2 Likes
I think you’re right that big boards aren’t supported
And no idea whats up with the top left!
2 Likes