Japanese rules multiply the effect of the technical sekis caused by open borders, since there are no points given to the entirety of the groups in seki. I think in Game 1, for example, White loses much less under Chinese rules. But almost everything on OGS is Japanese, so I’m going for Japanese here:
My algorithm would
not reveal AI invasions requiring the first move;
clean up trivial dame that are too small to have territory;
treat remaining open borders as seki;
not count territory for neighboring groups of a seki; and
take the Japanese definition of dead stones seriously (stones which can be killed in theory, not the ones that would be killed in one particular AI playout).
Game 1: I would mark all the white territory as seki, like @flovo, but I’d also give black his lower-left territory. So Black wins. (Should probably be White under Chinese.)
Game 2: White has no territory. Am I missing something?
Game 3: I’m not seeing any territory for anyone. White wins by komi.
Game 4: If White plays first he can kill all the Black stones on the right, so they are “dead” and should be removed. It doesn’t count as an “AI invasion” since it’s not penetrating a closed border. White wins.
Game 5: Black upper-right group is dead, and probably something at the bottom. But these are “AI invasions” that depend on moving first and thus should not be revealed. Black wins.
but auto-score does not. Easy for human to interpret, difficult for machine.
For example:
Where do you draw the line? Easy for a human, difficult for machine.
@gennan has already noticed a similar problem with allowing open borders:
An answer has been offered:
This is a good answer. Players should draw the line at the edge of their ability. But this is exactly the problem - it’s too subjective to program. The edge of the players’ abilities is different from the edge of mine or yours or @S_Alexander’s or KataGo’s. So players need to tell where the line is (by putting stones there, marking dead stones, etc)
All of these issues are exactly why we should have a dumb scoring algorithm. There are too many subtleties for a regular program to handle, and there is the issue of AI assistance when using a strong program like KataGo for scoring. In effect, the strength of the program needs to match the players exactly, which is impossible.
This whole discussion is silly anyway. Essentially, we are debating how to provide a feature for players to score unfinished games as if they were finished without closing borders, without marking dead stones, etc. It is going to be extremely difficult to program an auto-scorer that does this reliably as you basically have to predict where players will play. In fact, if there’s more than a handful of moves, then it is probably impossible to determine a unique answer due to the number of variations left. If it’s decisive, then those moves are needed to determine the winner. If the variations change the score but not the winner, then the loser should resign. If they don’t know whether it’s decisive, then they should play to find out. Allowing auto-score to determine if they should resign is external assistance. And again, where do you draw the line between finished enough and not? And all this just to save players from making the last few moves. What’s the point? Just keep it simple!
I think the sensible solution is:
Players close borders and agree on dead / alive stones, dumb algorithm counts points same as they would IRL
Or
Players leave open borders and leave without marking dead stones, dumb algorithm scores the game as is. If the players don’t like the answer, tough - should have finished the game.
Or
Players close borders but disagree on dead / alive stones, then life and death should be confirmed by resuming play (actual play for Chinese or hypothetical for Japanese) as per the rules. Then they can agree and dumb algorithm can count as point 1 above.
Or
In the unlikely event that players still cannot agree after resuming, then they should call a human moderator. This is rare enough that it is not burdensome on mods. Plus, using KataGo for this as an auto referee has problems as neither players nor KataGo can explain their interpretation of the position and KataGo may well look beyond the disputed local area to other parts of the board and thus inadvertently give external assistance.
As for this situation:
If it’s that clear cut, the loser should resign. If it’s close, then you need to finish and count anyway. If we run out of time at club, we record the game (phone photo usually) and finish it next time. Getting a program to count unfinished positions is like walking out of the club and asking the guy who stays behind to lock up to count for you.
If fully autoscore territory scoring the marked intersections with no user input whatsoever, I would given 1 point for the surrounded intersection on the lower side.
The new autoscore has become an issue because it is scoring boards on the basis of hypothetical variations it sees, but that the players did not see, rather than scoring it as it stands. This means, in effect, that KataGo has become a third player in the game. This has the consequence that in some cases the game play information contained in the KataGo variation can help one of the players to alter the game result, which is a violation of site rules against players receiving outside help.
I thought of a problem if we don’t auto-close borders. The losing player has no reason to help close them:
Why would White close the border? He will lose if he does, but if he passes, there’s a small chance that Black will also pass and everything will be in seki, so he can win on komi.
Instead he keeps passing, and Black has to play a bunch of extra moves to kill everything.
For the same reason White could just refuse to pass and play on until there are no moves left: it doesn’t increase their possibility to win, just prolongs the game. Black should respond by capturing everything that’s not alive and pass after there’s no dame left. Then there is no seki either.
But in case of playing on, mods can tell the player to just cut it out, pass, and accept the auto-score. Shouldn’t we want something like that to happen here?
Without knowing anything about scoring algorithms or reading the entire thread to find out, I would guess that it is an extremely difficult problem to solve. I would annull all games that can’t be sorted out by a simple “find crystal clear enclosures and ignore cutting points as well as whether groups are alive” algorithm, and let the players mark dead stones. Version 2 might attempt to automatically suggest dead stones using an algorithm that is as stupid as possible.
Can you really solve this in arbitrary games without playouts?
White’s strategy of trying to trick black into a technical seki is in violation of Japanese rules:
Therefore, white should lose by forfeit:
To any player faced with Black’s situation on OGS, I would recommend they first politely discuss the situation with their opponent in the chat to try to resolve it. And if that doesn’t work, then click “call moderator”
Ok, so let’s look at the question in a different light (which is actually the true situation in some of these games): one of the players was human, the other was one of the bots playing on OGS. Therefore, in the scoring stage, there is no “marking the stones” stage.
How would you solve this problem?
We can’t let the users mark the dead stones, since it makes cheating against bots very attractive, and hard to discover, since the bots likely won’t complain.
We can’t use your algorithm, since it generally will mark most of the board as seki as soon as there is some dead stones left on the board.
We also can’t require the players to capture the dead stones before passing: even if we could human players to do so, how are we ever going to convince amybot?
What if the bot player is the one who doesn’t close their border?
What if both players are bots?
Naturally ranked games against bots can be disabled, but that doesn’t really solve the problem, since unranked games need to be scored as well.
KataGo could be that third party, but so could any other bot that can tell the life-and-death-status of stones.
Or human volunteers could do it, but without a ruleset to decide, different volunteers would give different outcomes. And if there was a ruleset, it probably could be turned into an algorithm.
Why does OGS find scoring so difficult and generate endless forum threads about it? KGS doesn’t.
To answer the question of this thread. Easy. There is no auto score. Scoring is part of the game (in which you mark dead stones and territories), if players leave without scoring then the game is simply unfinished. Not annulled. Not scored. Not by a superhuman AI or a simpler algorithm.
