I recently finished a go game with GnuGo, and somehow the system turn one of my very alive stone dead, and then declare the entire circled area as non-territory even though it’s my territory.
As the result, I lost the game even though I should have won.
Please review the game again for me as this affect my rank quite a lot when OGS keeps having these miscalculation errors left and right.
Here is the link to my game:
Sorry about that. It is sometimes a bug that happens where games against bots are automatically scored incorrectly.
We have annulled the game.
Fun fact: part of the chat is automatically translated.
This is a well-known problem. Two threads have discussed it. One was closed. One is still open: A compendium of OGS's terrible scoring system confusing beginners. The problem is that the AI interpolates its own judgment rather than scoring the board as it stands. The reason the whole corner is unscored is that the AI thinks the border is not technically closed. You are right that the game should be annulled (game results cannot be reversed). You can file a report from the game tab (click the triangle on the right-hand panel), and a mod will annul it.
Edit: I am now late, I see. I write so slowly.
This kind of thing happens frequently in non-bot games as well, as the previously cited thread shows.
in non bot games its possible to choose status of groups
Lots of people, even a few experienced players, don’t know how to correct the scoring. Hardly anyone reads the documentation.
if page design would be changed, no one would need documentation
Future comment by Conrad:
Hardly anyone reads the page design.
Not my bailiwick. You’re talking to the wrong person, and none of this changes the truth of what I said.
In this game i can’t guess what is confusing the scorer. Weird.
I think this blow up of the down right corner explains it all.
As you can see the black stone at T5 is transparent.
If you hover with your mouse over the board , it looks like this.
It seems that Black did hover there, but didn’t click (to place the stone there).
The bot doesn’t see that stone that isn’t there and decides that the right bottom area is still open. And does’t count it as black territory.
Are you sure?
I took the final board position of the game that @mimiucute1509 linked in the OP.
So I think the bot assumes that move 147 was not placed there.
Something is wrong with that bot.
By your theory bot thinks that this ghost stone was not placed too?
The bot opponent does not do the scoring. Auto-scoring of bot games is handled by a different agent, which had an evaluation bug.
It seems like if we defer to kata to determine dead stones, it might make sense to get the score/area map from kata too. Even with that stone dead, the larger group could have lived
This is the most surprising insecure corner I have seen!
Autoscore v3 shows some hint about what’s going on:
With White to play (I guess all bot games are scored this way?) KataGo thinks White can live in the corner. Does it end up as a seki? I’m still not sure how that (gray) stone is supposed to end up dead.
(Of course the autoscore v3 algorithm would get it right, leaving all those stones alive since none of their statuses were ever in question.)
Time to abandon this theory I think.
Still curious what caused this.
Are there good reasons at this point to not try switching to the “autoscore v3” algorithm as a replacement for the current algorithm?