The mistake is the players. Games need to be properly completed to be properly scored. The ‘resume’ option is available to fix a situation like this for scoring.
3 Likes
Why? There aren’t any dead stones there so the four empty spaces should be dame.
1 Like
This position is considered settled. Black doesn’t have a move to make it into territory and white doesn’t have a move to take more points than just that one here. If it appeared in an online pro game they would score the game without adding any moves here. These kinds of endgame positions are very common.
2 Likes
Of course. They would score it by agreement and indeed the scoring system allows users to do it that way manually if they wish but it’s not the role of the scoring algorithm to make assumptions, even valid ones, about what’s what with endgame.
[Edit]: Game link please!
1 Like
There is no such thing as not properly completed.
Any game is completed if the players decide to end it here, and any such game can be scored. Yes, if there was a mistake such as forgetting to close a territory then the score might be quite different from what the players expected, but this should not concern the scoring system.
Here the issue is that the scoring system scored the game wrong.
6 Likes
To correct you have to resume the game and play the dame. It’s not like you can fix it by telling this or that stone is dead. If we used chinese rules that would be fine but we are using japanese rules so we shouldn’t have to finish these dame points.
1 Like
The game-log is odd and hard to read. As far as I can tell (and I’m not certain), the auto-score failed to operate on the first double pass, you attempted manual scoring in a number of steps before rejecting that and resuming (a good move) and then comes the uncertainty because it looks like the auto-score operated the second time scoring started but there is a line or step missing from what I would expect to see. Sorry I can’t be more help.
What I can say is that when the auto-score completely fails to operate, what you will see is no stones marked dead and territory only counted if it is devoid of opposing stones (because they have been left marked alive). The easiest solution when the auto-score fails to operate is to do as you did and resume the game, then reenter scoring which kicks the system into another attempt.
As for the scoring at J18, I’m perplexed. Typically the “open border” would prevent such a point being counted by our system but I am used to Chinese, not Japanese scoring in my games. However, given that users have the option of resuming games I still think it is hazardous to have the auto-score make any assumptions about endgame.
1 Like
Well yes i resumed to correct other things and then didn’t have enough scrutinity about this,last one. Didn’t change the outcome and i am used to play the dame anyway, won’t try again japanese scoring
2 Likes
As often happens with the passage of time, there seems to be widespread forgetfulness about previous discussions on this subject. See Autoscore is still overriding players’ skill level
The current autoscore remains defective in at least two ways. The autoscore bug—when closed territory is left unmarked—which began roughly in August of 2020, still exists. It was greatly reduced by the last autoscore revision, but still pops up occasionally (e.g., sev ra vs. Fard, or Zeraora vs. GeorgySmolyar).
More serious is the issue that was the focus of the cited thread, and which still occurs now and then. That is, where the autoscore scores the game based on an AI variation rather than on the basis of the board as it stands. Recent examples: jaelgraham51 vs. KingDe3, Partida amistosa. This is a violation of site rules because it constitutes AI assistance to the player who shouldn’t have won (because he didn’t know the tesuji). Until this is fixed, it means that a player not only must beat his or her opponent, but must beat the AI autoscore.
8 Likes
2 Likes
To summarise the problem here.
When 2 humans score a game they can mark dead stones, come to an agreement and score that board position.
When a human plays a bot OGS currently uses the following protocol: a superhuman AI is used to mark dead stones, not based on the rules of scoring, or the players agreement, but based on continued high level play. This can result in invalid results that violate the scoring rules of Go and marking things differently to the players would where it spots weaknesses the players didn’t. Admins sometimes intervene to annul them.
An alternative approach would be to let the human mark the stones and the bot offers no opinion. I believe this was used in the past. A problem with this is some humans are naughty and mismark dead stones to get a win they shouldn’t. When this happens admins could intervene and annul/warn/ban. I don’t know the relative incidence of naughty humans vs naughty AI autoscore.
A better alternative approach would be to allow the bot to offer its opinion of dead stones, just like a human, and only score on agreement. This is technically possible in the gtp protocol, hexahedron has detailed it previously. KGS does this for example.
5 Likes
I think we could in theory intervene but some bots play like thousands of games in short spans of time, and probably less people would be watching bot games at various levels in order to report any naughtiness. So I imagine quite a lot of it going unnoticed 
People are naughty in other ways with bots too, like if they can’t read ladders or make the same joseki mistake each time in a given situation.
But one could also think of making games unranked only for it being somewhat useful for new players to get a rank (that includes beginners and dan players beating something like doge bot etc).
I had thought about whether one could have ranked games against bot not affect the bot rank but only the humans to combat some naughtiness but I’m not sure it works
It would be better definitely, but possibly more work for the admins of each bot if they have a somewhat custom bot. It might be zero extra work for some though.
There’s definitely been some good ideas to improve scoring also from @Vsotvep , @MrAlex , @Feijoa etc I believe, in particular with beginners in mind, but it might take some work and consensus.
3 Likes
4 Likes
I can’t tell how that situation came to be.
It’s possible that white correctly marked those black stones dead and black accepted. In that case, there was no fault made by either player or OGS’ scoring system, and it 's just a case of a beginner being confused by the correct result.
3 Likes
Uberdude’s example is such that
there is an unambiguously correct scoring
and
the shown scoring is that scoring (since Black is dead as it stands on top)
, so if there was no manual marking, then this is a
clear-cut case in which OGS’s scoring system was correct.
(So, it’s an example for “confusing beginners” ,
but far from being an example for “terrible” .)
4 Likes
“OGS’s terrible scoring system” was correct? Is this a problem?
2 Likes
1 Like
For that one, my view is that F1+F2 and the rest of the right should’ve been scored as Black’s:
This is area scoring, so not Japanese rules, and Black-to-move easily secures F1+F2 and the right,
so the corner being unsettled should result in falling back on Tromp-Taylor for the right.
(Everyone here should agree that if there was still play going on elsewhere, then those stones
would be unsettled rather than dead, and unlike technical Japanese rules, I don’t know of any
area scoring rules that would at-all-clearly score such stones as dead after consecutive passes.)
I don’t know autoscore’s details, but I imagine that
Black passed last and
autoscore accordingly only considered White-to-move
.
If that’s correct, then the following is my suggestion:
Run the AI score estimator twice, once for White to move and
once for Black to move. If those 2 runs agree, then use that, else
treat only the agreed-dead stones as dead, and otherwise apply Tromp-Taylor.
2 Likes
Tromp Taylor sounds very tedious for bot vs human games, which I believe is where there’s a number of autoscore issues. The bots can’t argue the score, and the human doesn’t want to have to capture every dead stone on the board.
But they could, and do on other servers. The gtp protocol includes bots marking dead stones in scoring phase. OGS has not implemented it, choosing to use AI scorer instead as a workaround. A demonstrably poor choice.
1 Like