Here’s a question related to improving the (weak, non-AI) score estimator, which only knows about area counting. Consider a 5x5 game with no komi and no captures yet:
Given this prediction from the Score Estimator, what should the score be under Japanese rules?
B+3.0
B+2.0
B+1.0
W+0.0 / jigo
W+1.0
Impossible to determine
Other
0voters
Note that normally the Chinese and Japanese scores differ by at most a point, but someone might have passed early, and the game is not over. I know it’s not very complicated, but since we’re talking about Japanese rules, there might be something to debate even with this one!
If you can think of any stranger cases for the non-AI score estimator, please feel free to share below.
I think that the weak SE has done his job already and that it doesn’t really matter furthermore. He got that both groups have two eyes. It’s used ingame so that’s enough infos.
So i voted others
Under Japanese rules, the SE should not put marks on stones, except when it thinks they’re dead (as E4 in the diagram).
Then I see 5 white points and 6 black points. There is no komi, so black is leading by 1 point would be consistent with the territory markings in the diagram.
Edit: I think those territory markings would be wrong for territory scoring. As the SE uses area scoring under the hood, what would the markings at D1 and E1 mean? If it expects stones there, D1 an E1 should not be marked at all, because those will be dame points or boundaries under territory scoring.
Edit2: The difference between area scoring and territory scoring should be small in most cases and the SE is purposefully weak and known to be inaccurate. Why bother to tweak it for territory scoring?
There should be “estimator” where you can freely and easy paint like in Paint. And this estimator would just exactly count area of what you painted. Nothing Go related.
Something like Voronoi_diagram(but with low resolution) may do default paint (before you repaint it).
Drawing lines would be great, but this improvement is much easier:
If we just subtract all the “own” stones and dead stones (and add any prisoners - thanks @benjito), we arrive at something that looks like a Japanese score.
Yes, unfortunately it doesn’t reflect any possible continuation, especially since there’s no way for black to hold the territory at A4. But maybe that’s the kind of mistake you should expect from the weak SE anyway?
Maybe it’s easy enough to correct the SE results and stone markings for those specific cases, after the SE did its thing, instead of branching the SE into 2 versions?
That’s what I’m doing, just modifying its area-scoring results to look like territory scoring. I don’t want to even look at what the SE is doing internally, it’s too scary
This is a complicated question, since there are two different score estimators and different contexts (which rules, whether the perspective is a player or spectator) to consider.
My understanding (and I might be wrong about the exact specifics), is that two score estimators are as follows:
A weak score estimator available to the players of an ongoing game (provided that analysis is enabled). This weak estimator only provides an area score estimate, and is known to be highly error prone, often making misjudgements about life and death and will often wildly fluctuate between separate queries on the same position. This one was developed long before the AlphaGo era and the availability of superhuman Go AI engines.
A stronger score estimator that makes use of a strong AI engine. This one is only available to the players after the game has finished, but it is always available to the spectators. This engine is fairly accurate and will appropriately provide either a territory or area scoring estimate depending on which rules the game is using.
There are at least three more score estimators on OGS:
The scoring tool you access by clicking in the player names / captures / time section at the top of a game. This one is very accurate but has no intelligence - you have to close all borders and kill all dead groups to use it. Also it doesn’t actually add up the points for you.
The game-end autoscoring tool. This understands rules and dead groups, and unlike the AI score estimator, it makes a clear decision about the status of every point, to always return a valid score (i.e. ending in .5 for normal komi).
The AI analysis graphs generated after a game, which are often quite different from what I see in #2.
I was hoping for some unusual cases in Japanese rules to come up for discussion, but maybe the weak Estimator is too weak for that anyway; it doesn’t even know about atari.
