For the last several months, KataGo has been computing a 9x9 opening book. Today I decided it was big enough, and posted it online at https://katagobooks.org/ (download links are also here if you want the book for offline).
There is both a Japanese rules 6 komi book and a Tromp-Taylor rules 7 komi book. Direct links: 9x9 JP book 9x9 TT book
One fun thing is that 3-4 appears to be likely suboptimal as the first move! This contradicts some prior KataGo opinions that had it as likely optimal but narrow/difficult to play. But of course, maybe a deeper analysis in the future will contradict this one, or someone will find a key blind spot somewhere.
Also the page has a small showcase of fun and weird 9x9 tesujis and such, a few of which I’ve posted separately about on this forum, but there are some new ones too.
Thank you for sharing KataGo’s new 9x9 opening book! As a 9x9 enthusiast, I really appreciate the amount of effort and computation that went into generating this resource. I also appreciate the shout-out in the overview.
That 3-4 may be suboptimal as the first move is fascinating. Move 10 at E7 wasn’t new to me, but KataGo’s deeper analysis revealed the excellent A2/F5 combination I had missed. The highlighted unusual tesujis are also a delight. Such creative play. I’m looking forward to studying these books.
KataGo keeps pushing the boundaries of Go AI. The fine-tuned 9x9 net looks impressive. Releasing these books and nets for free is so generous. Opening up this knowledge advances the whole Go community. Thank you again for this wonderful gift! I’m sure many players worldwide will get great benefit and enjoyment from it.
Awesome—you beat me to it: I just came here to share this after reading it in the “Computer Go Community” Discord server, and I’m glad that I searched for “katagobooks” before actually sharing it
[A few seconds later]
Oh, YOU decided? So you are David J Wu (“lightvector”) himself, the developer of KataGo? What an honour to have you here! Thanks so much for your work, for making the powers of Go AI accessible to all of us!
Also for those who might not have drilled down to this point, about the special 9x9 net:
One fun feature is that this net also has a little bit of training for 9x9 handicap games, including the “game” where white has a 78.5 komi** while black has 4 or 5 handicap stones, such that white wins if they live basically anywhere. This training did not reach convergence, but enough that if you try searching with a few millions of playouts, the results are pretty suggestive that white can live if black starts with all four 3-3 points, but not if black gets a fifth stone anywhere reasonable.
(**Area scoring, with 0 bonus for handicap stones as in New Zealand or Tromp-Taylor rules. If you use Chinese rules, you’ll need a lower komi due to the extra compensation of N points for N handicap stones, and if you use Japanese rules you’ll need a lower komi since the black stones themselves occupy space and reduce the territory. Also leaving a buffer of a few points from 9x9 = 81, like choosing 78.5 instead of 80 or 80.5 is a good idea so that the net is solidly in “if I make a living group I win” and is well separated away from “actually I always win even if I lose the whole board”)
I haven’t looked at things like this in detail and analyzed how the variations actually work though. I’ll leave it for anyone else who is interested to explore for themselves, and/or post if they want. Also I think it shouldn’t need quite as many playouts after finding the main variations and “helping” things along by playing out some of the moves, the millions is just from the opening position.
Well, that’s amazing in and of itself, but it just so happens that I was thinking this weekend to get the board out of the storage and resume studying Go after all those years. And here you are posting this and another fellow in the FB group asking for “where to find pro games to review on a real board”.
People worry about ad algorithms “reading their minds”, but happenstance does that too. Thank you very much for your effort and for sharing this
