Katago vs score difference

General Go Discussion
1

The game result is a W+0.5. Chinese rules, 0.5 komi, 1 handicap. Katago, both from OGS and a local view in katrain suggest at the end that black is winning. Why?

I’m sure there is an easy answer related to the 1 handicap, or the prisoner at M18 not getting counted correctly.

image
image1920×1131 226 KB

The prediction changes once I just place the black stones like this (M19, N18):

image
image1920×1131 229 KB

What I don’t understand is how the predictions can be off in such a critical point where I’m sure that they generally support Chinese rules …

Thanks for any help :slight_smile:

*Update: added missing game link.

2

[Deleted by author]

3

The fast review can be mistaken on the score estimation, if there’s any ataris and things that don’t work but go a few moves deep.

So my guess is that it could be to do with katago’s playouts maybe?

Edit:

I think I counted twice and got 181 for black. So that should 180 for white, or B+0.5?

image
image1420×1422 227 KB

image
image1426×1426 259 KB
n

So then did the scoring go weird?

Edit 2:

image
image974×406 47.1 KB

Or if you add a few extra black stones:

image
image974×412 47.2 KB

So actually it’s behaving more like a komi=1.5 game then if it’s awarding White the handicap point. I don’t know if that’s intended for 1 handicap.

1 Like
4

In Chinese Rules, white gets one extra point of compensation for each handicap stone:
OGS rules comparison matrix

KataGo supports this (KataGo rules, see “WhiteHandicapBonus”) but the OGS KataGo might be configured incorrectly.

1 Like
5

For each handicap stone yes, but I don’t think it’s supposed to be for handicap 1, which is no extra stones, just reduced komi.

image

But I guess it’s been like that for a long enough time?

1 Like
6

Yeah the “N-1” option (as used by AGA rules) makes more sense. In any case, there seems to be a mismatch between the implemented rules for the scoring and for the AI evaluation!

4 Likes