Preliminary analysis of 9x9 openings with KataGo

For discussing go computing, software announcements, etc.
Post Reply
iopq
Dies with sente
Posts: 113
Joined: Wed Feb 27, 2019 11:19 am
Rank: 1d
GD Posts: 0
Universal go server handle: iopq
Has thanked: 11 times
Been thanked: 27 times

Preliminary analysis of 9x9 openings with KataGo

Post by iopq »

I didn't put in a ton of time on this, but at move 1 it already seems basically correct

This is Japanese KataGo rules with 6 komi using a 60b network

To have actual good analysis I would need about this many playouts on each variation, so I'm still ways out from making any actual claims other than it looks like there are four moves that are equally good in the opening (I verified by testing the 3-3 and 3-4 variations manually)
Attachments
Screenshot from 2021-10-15 23-26-31.png
Screenshot from 2021-10-15 23-26-31.png (171.28 KiB) Viewed 5798 times
iopq
Dies with sente
Posts: 113
Joined: Wed Feb 27, 2019 11:19 am
Rank: 1d
GD Posts: 0
Universal go server handle: iopq
Has thanked: 11 times
Been thanked: 27 times

Re: Preliminary analysis of 9x9 openings with KataGo

Post by iopq »

I've already verified that 3-3 and 3-4 are not optimal by searching from those positions manually

So the top four moves are:

5-5 tengen
4-4 hoshi
5-4 takamoku
5-3 mokuhazushi
Attachments
Screenshot from 2021-10-31 17-40-08.png
Screenshot from 2021-10-31 17-40-08.png (166.24 KiB) Viewed 5481 times
Screenshot from 2021-10-31 17-39-54.png
Screenshot from 2021-10-31 17-39-54.png (167.85 KiB) Viewed 5481 times
Screenshot from 2021-10-31 17-32-37.png
Screenshot from 2021-10-31 17-32-37.png (164.7 KiB) Viewed 5481 times
iopq
Dies with sente
Posts: 113
Joined: Wed Feb 27, 2019 11:19 am
Rank: 1d
GD Posts: 0
Universal go server handle: iopq
Has thanked: 11 times
Been thanked: 27 times

Re: Preliminary analysis of 9x9 openings with KataGo

Post by iopq »

I'll get a bit more time on it to post the full HTML files, but you get the idea

Most lines close to 50% are most likely to be optimal, lines outside +/- 5% are more likely to be suboptimal
Attachments
Screenshot from 2021-10-31 17-40-08.png
Screenshot from 2021-10-31 17-40-08.png (166.24 KiB) Viewed 5481 times
Post Reply