KataGo Distributed Training and new networks

[And]

BadukAI s555 1p - Sabaki b40c384 1p 1:0

[go4thewin]

katago 40b s558 1po vs golaxy 9d : 1-0
b+r
http://go.ba.net/playgo/go-embed.html?s ... f&move=151

katago 40b s5675 1po vs KGS Hirabot 6d (10,000 po?)
b+r
https://kifu.io/view/t3ArnpUDdJDzfWjdU4x2

[And]

KataGo started winning with black. BadukAI s560 1p - CS Zero 9d 1:0
after 231 moves, CS Zero showed Black's winrate 12.5

[And]

CS Zero has no chance against the new s567 network! KataGo s567 1p - CS Zero 9d:

[And]

KataGo s572 playouts 1, komi 0.5 - CS Zero 9d:

[go4thewin]

I hope s572 or s570 is the net that gets optimized. It's strong enough now that no/few users will have to play against more than one playout. Great for a mobile app!

Just like against cs zero 9d, It also beat the strongest 7d rank on zen 6 (gui version, I have not tried gtp4zen). Using 1 playout, playing as white with no komi, zen resigned after 200 moves; s572 was leading by >70. zen 6 is a strong program that many dan players love to play against.
http://go.ba.net/playgo/go.html?sgf=3mDscV5Cq.sgf
https://ibb.co/GFTLG7j

[And]

KataGo s572 playouts 1 - Zenith 7 9d:

[And]

with the same time settings, CS Zero plays stronger
CS Zero 10 sec - Zenith 7 10 sec 14:6
CS Zero 60 sec - Zenith 7 60 sec 12:8
but when playing 9d it seemed to me that Zenith 7 is stronger. KataGo s572 playouts 1 without Komi could not win against Zenith 7

[go4thewin]

thanks for this, i was curious about it!

[And]

KataGo s572 playouts 1 - LZ #286 playouts 5, LZ surrenders on move 171

[lightvector]

New distributed training run should be now open for anyone to connect!

viewtopic.php?f=9&t=18019
https://www.reddit.com/r/baduk/comments ... ributions/

[wineandgolover]

Hi lightvector,

Are any of your contributors using Mac? I’m happy to help, but I'm probably a bad choice for a pioneer. My GPU's aren’t bad though.

[hakuseki]

This is just an idle thought, but would it make sense to increase the number of visits in self-play games in the later stages of a training run (i.e. now)?

My thought is that reinforcement learning will converge to a plateau where the approximation loss (e.g. the trained policy is 400 elo points weaker than the player it is approximating) equals the gain from tree search (e.g. 600 visits is 400 elo points stronger than 1 visit).

To move past the plateau it's necessary to either improve the approximator (e.g. by increasing network size) or approximate a stronger player (e.g. by increasing visits in search). I know KataGo already uses the former, but what about the latter?

[lightvector]

Hi lightvector,

Are any of your contributors using Mac? I’m happy to help, but I'm probably a bad choice for a pioneer. My GPU's aren’t bad though.

Unfortunately I don't own a Mac, and don't have a great way to compile and test myself. I expect probably some people are using Mac for this, but if so, it's mainly going to be through building from source on their own. Which probably is not *too* hard, since (unlike Windows) I think package management and installation of libraries on Mac is pretty easy, similar to Linux? Things are a bit wordy here: https://github.com/lightvector/KataGo/b ... mpiling.md but the actual process is not hard (except on Windows).

Right now the release process is also a bit time-consuming for me, to compile everything on my own personal computers (windows and linux) and package stuff up. If someone were willing to put in the effort to help set up a more automated build process for releases, which could maybe include Mac as well, that would be fantastic.

[And]

KataGo s581 playouts 1, komi 0 - Zenith 7 9d: