Rare result

Made me laugh, not sure if it can be described as a bug, but it certainly isn’t right :sweat_smile:


image

1 Like

Do you have a link to the game?

By the way, it is possible to set a custom komi value, like 6.6, which doesn’t really change the game.

With uploaded SGFs, one could have all sorts of interesting results

1 Like

Oh sorry I didn’t include the score at the top, it was a .5 Komi

1 Like

That game ended with a result of black wins by resignation.

The screen that you shared seems to be from the score estimator output, which does not round the estimated score to the nearest half-integer, and displays values like that to express a degree of uncertainty.

5 Likes

Ah Ok, understood thanks!

Percentages and point difference are very different. I’m not surprised that the AI thinks that a 1.5 point lead when the game is over corresponds to a 99.9% winrate.

Or is that not the “bug” that you wanted to show?

(BTW, I think point difference is much more useful for humans than winning %)

1 Like

I was under the impression that the supposed bug was simply that the margin was 1.6 points instead of 1.5 points.

3 Likes

It’s often surprising how uncertain the strong post-game) score estimator can be even in the very late endgame:

(Move 27 from this game)

1 Like

what’s uncertain about this? it accurately predicts that at this point the game was even…

1 Like

I think when we can accept things like B+1.6 I think we also have to accept W+0 etc.

I suppose there’s a simple continuation where Black wins by half point but there’s other trades to try etc that might muddy the score estimation/averaging.

I don’t think the score estimation tool is necessarily coupled to the ai review, and I’m not sure what the equivalent level is for example. Rerunning it a few times does give small changes like B+0.1/0.2/0.3 etc.

I do feel like mentally rounding to the nearest 0.5 point score could work in cases where it hasn’t settled on an exact midpoint between the two.

1 Like

Is there a continuation that forces a draw? If not, it can’t actually be an even game, and what’s surprising is that strong AI can’t just completely solve it at this point. Of course, I might be wrong and it might be seeing something that’s worth a try for white, but I suspect it’s more that it has a bit of a weakness in evaluating endgames.

Also, you can try evaluating all the possible responses by White. The obvious one evaluates to a more reasonable B+0.4, and there doesn’t seem to be anything better.

And oh, looks like someone ran a level IV review for me. (Thanks!) It shows 93.6% win rate for Black and suggests playing out the ladder.