On some borderline cases where there’s a lot of dead stones and the margin of victory is huge, the score estimator seems to be way off even at the very end of the game. At worst cases i managed to find its almost 20 points off, and i can’t see any reason why this happens.
few examples
Granted, in games like those no-one really needs the SE to see who’s winning, so its prolly not that big of deal, but it’s baffling how the estimator can be so much off. Neither of those games had sekis or anything else which could explain the difference
Ahh right indeed!
i somehow missed that and only thought about the difference of captures + dead stones xD
But i dont get why is that messing with the estimator. The value of captured stone in japanese rules versus the value of living stone in chinese rules is the same, so the margin shouldn’t change, right…?
I will say that I do quite like the score estimation as it is, drawing the squares on the board as a toggle on and off, specifically after a game is finished.
It was one thing I though was quite cool that AI sensei had in their reviews, that one could visualise the territory on the board.
I also prefer the way OGS puts the score/% loss on the moves rather than AI senseis A,B,C… etc but now OGS also has a way to display the counting on the board, which I think is a nice touch for reviewing.
This is a great point. The auto-score algorithm should not depend on whose “turn” it is, right? And now it does?
Even with all the unsettled groups handled, there’s still the possibility that auto-score will reveal completely dead groups that neither player would have noticed. Obviously there could just be dead shapes around, but another way this could happen is that with a few subtle weaknesses on the board, if some of them can be killed in sente, some will be dead no matter whose “turn” it is.
It will be interesting to see what strange cases arise in scoring.
It should definitely depend on whose turn it is, like in your game, if it’s white’s turn, the estimator is correct in that the top part is not territory. There’s plenty of places where life / death depends on who plays first.
However, when the score estimator is estimating the score, it shouldn’t give an estimate for unsettled territory (it’s unsettled after all). It should do that at the end of the game, which is what your problem was.
I mean to say, the score estimator should use the information of whose turn it is to make sure the score doesn’t flip back and forth depending on whose turn it is.
In a capturing race, the SE should ideally be able to know who will win, right? It shouldn’t flip back and forth depending on who had the last move.
(apart from whether this is something we want of a score estimator, since it may be too powerful)
I was trying to say that at the end of the game, after both players pass, “auto score” should not be using information about who passed last, since we’re done with turns.
By my logic, the estimated score depends on whose turn it is if and only if there is unsettled territory. So if SE were to not estimate unsettled territory, then it would not depend on whose turn it is.
I agree, but it should evaluate the game as if black can play another move, and as if white can play another move. If the game is properly settled, these should be equal. Any discrepancy between those two is unsettled area, and should be dealt with, for example, as I suggested above. Right now it seems to just choose the position where the player who passed first gets to make another move.
Yes, but we’re talking about two things here: SE during the game and SE to determine the score at the end. For the first unsettled territory shouldn’t be marked, for the latter the SE should mark it in a way that seems reasonable to both players, even if the AI considers the game to be unsettled.
I think unsettled area during the game is a bit vague, and it’s probably fine if an estimator can give some points in areas that are a little uncertain. (Of course you don’t want that if you’re using this kind of thing in ranked games where you don’t want a strong ai opinion of what will likely be your territory.)
On the other hand at the end of the game “what seems reasonable to both players” is also a bit vague but I do have a suggestion:
Let Katago decide what stones are alive and dead based on whether they are “pass-alive”, that is if their status hasn’t changed with the last two passes. Any remaining groups whos status has changed with the last two passes should be marked alive contrary to what katago would do depending on whose turn it is. That way it hopefully should reflect what the players believe since they both passed.
How about this: If both players pass and the AI concludes that the game is not finished and the estimated result of the game depends on who plays first, then the game is declared a draw.
