I just uploaded a 9x9 game i played the other day to have it reviewed by KataGo. Since it happened on a real board, i had to replay it from memory in an SGF-editor (I used Sabaki).
The analysis offered seems quite off however and i was wondering, why this is. Is it because its a 9x9 game or because of some other reason, for example that it’s not a “real” game file, but rather a replay?
Das Spiel ist beendet (online-go.com)
The analysis should work quite well for 9x9 games, I suspect the issue must lie elsewhere. Would you be willing to share the original sgf file?
I don’t see the same thing.
Yeah, sorry that’s my bad. I triggered a more detailed analysis to see what happens. You can view the original one by clicking above the graph.
It still seems weird though that it would go for such weird moves even with less playouts
Oh I see, analysis level II doesn’t work so well on 9x9, probably because Katago was trained on 19x19.
9x9.sgf (347 Bytes)
Here is the SGF-file I uploaded.
@jlt Hmmm … If thats actually the reason, I find it a bit weird. You might as well make 9x9 analysis availible for higher tiers only, if it shows nonsensical moves like this one
@kickaha , this bug seems new and happens on 19x19 too
Okay, I see. In that case ill try to upload a few more. Maybe i can figure out how to reproduce it.
I don’t think that’s true, see [1902.10565] Accelerating Self-Play Learning in Go
And 7x7 and 8x8, in the years following that paper.
Which is why we can now “solve” 7x7 and 8x8 better than any prior attempts at determining the best play on those board sizes (https://katagobooks.org/).
Updated the other thread with some more unlikely speculation. AI suggests nonsense moves on 1st line - #6 by hexahedron
I guess only anoek or someone with more detailed access will ultimately be able to say.