Yeah I guess so, the bent four is killable, but then the autoscore marked the seki.
Maybe one another case for improving the autoscore then
Yeah I guess so, the bent four is killable, but then the autoscore marked the seki.
Maybe one another case for improving the autoscore then
It’s interesting how the autoscore recognised the bent four is dead but didn’t recognise that the seki is fake. I wonder what’s the logic.
Can you really expect the autoscore algorithm to correctly score positions like this?
We can construct arbitrarily complex positions, you gotta draw the line somewhere.
If humans can evaluate it, I imagine we can expect AI to figure it out at some point?
At anyrate, the complex positions make for good training data
Katago could probably score it correctly, but then we added logic on top of that, because Katago wasn’t scoring like humans that couldn’t spot weaknesses in positions like double ataris and more complicated tactics, or just generally unsettled positions like
yeah, I think this is fine by autoscore; we don’t want it playing for people
Is this the same autoscore that scores bot games at game end? Should that really be nerfed?
If you’re playing a strong bot maybe it doesn’t matter if it’s a katago scoring it. If you’re playing a 25kyu bot maybe it matters.
But the scoring here just doesn’t make any sense. If the corner is dead as shown, there is no seki.
If the goal of the autoscore is to score a game as the players would, it’s not achieving it.
it’s arguably unsettled, so should we really expect well-defined behavior? this is a good deal less clear than just a missed cutting point which the autoscore should obviously ignore
I would not call this position “unsettled”, as the situation does not change if either player passes and let’s the other play first. Among two experienced human players, it would be reasonable for them to both pass in this situation and accept that the 15 White stones in the top-right are all dead.
At most, we could maybe say that the situation might not be clear to less experienced players, who might misplay while trying to resolve a life and death dispute. Among such players, it could be reasonable for White to dispute the life and death status, forcing Black to play it out and try to prove this (of course, this is in the context of Chinese or other area scoring rules), since Black could possibly make a mistake.
As this is a bot game, we could, of course, say that it would be prudent for Black to play this out, in order to ensure that the auto-scoring does not get confused like this. However, that’s assuming that Black is the human player and is aware of these potential quirks of the forced auto-scoring for bot games. On the other hand, what if White was the human player and Black was the bot? A bot might properly recognize the situation, but not necessarily be programmed to clarify it with further playout, before passing. If the bot passes as Black in the above situation, should a human player as White be compelled to further playout the situation just to clarify to the auto-scoring that their own stones are dead? I would think that a human playing as White, if they cared about getting the game scored correctly (even possibly to their own disadvantage), should do so, but it would be reasonable for a typical OGS user (unaware of such bot-game auto-scoring issues) to not do so.
Can someone post a link to the game? Edit: see my reply in the other thread.
I agree, except that while I think it’s reasonable to expect a human player to resolve situations to their advantage, I think it’s the bot’s fault if it doesn’t do the same and it’s perfectly legitimate for a human to take advantage of that
2 posts were merged into an existing topic: Auto-score improvements
Precisely. If the algorithm thinks that the corner is dead, then it doesn’t make sense to mark it as a seki, so the logic is not adding up. I’m not sure whether it’s bent four that’s making it complicated. Probably we need another example where white is dead but not in bent four.
I suppose in Chinese rules the stones are worth points.
In Japanese rules, capturing the single stones in the bottom left and top right are worth points, but cancel out.
So in Japanese rules whoever has the most prisoners, adjusted for komi, wins