Is this broken or are my eyes broken?


Why is it B+1.4 in the black bar and not B+1.5?


Someone with a better understanding of the AI than I can explain the mechanics behind this - but it’s enitrely normal for the AI to estimate in .1 increments, not increments of 1.


Reminds me of how one server specifies the komi tiebreaker in its SGFs as “.5000000” . Obviously any number 0 < x < 1 does exactly the same job.

It’s actually been suggested before that the idea of tiebreaking being represented as a part of komi is flawed, and that it should be explained separately. I think this is what the Ing Cup does.

A format like W+6&T (“tiebreaker”) could also be used.


It’s the final score, why are they different?

I can positively say I understood nothing. :frowning:

1 Like

The reason for this is that the AI is quite the optimist. It sees a tiny tiny chance for white to gain a few points, even though that will never happen under reasonable play. It then factors that tiny chance into the average, so it says that it will be Black by 1.4, not 1.5.

If the difference is much higher, like 10.9, it means the AI probably sees something we don’t, but as long as it’s ± 0.5 and there are no kos, it’s safe to say the game is over.

You can see if the AI has any good ideas by checking if it has any moves on the board that significantly affect the score. Anything below 0.2 or -0.2 is probably gibberish though.


What I’m saying is that as a tiebreaker, 0.2, 0.3, 0.4 etc. do exactly the same thing as 0.5 when all’s said and done.

Fractional points are only used in endgame theory and represent the chance that a player will be able to make certain moves – they’re an approximation tool.


What is the fuss about .1000000 points?

@Atorrante I’m not sure I understand your question.

Why make a problem of .1 point?
Fuss is a display of unnecessary or excessive excitement, activity, or interest.

Why reply to a question you haven’t understood?
Maybe don’t reply to everything just to reply?

It works the same as chess AI scoring in 1/100ths of a pawn… AI need accuracy greater than integer level when making predictions, even if ultimately those numbers can’t end up being a real score


I would say that AI always estimates outcome and it does with “floating point” numbers, which means that there are many more digits that simply aren’t displayed.
That “1.4” could be a short way to say 1.44963647073268 or something.

While the final score is necessarily integers + komi and that’s where the .5 comes from.

I don’t think AI ever estimates score by integers.