Today’s update brings several improvements to the AI system:
We now support both KataGo and Leela Zero engines.
KataGo is now the default review engine, Leela Zero is available for 19x19 games by clicking the review list drop down and clicking the Leela Zero button.
9x9 and 13x13 games are now supported (by KataGo only)
We’ve dropped support for the 15x128 Leela Zero network and now just use the 40x256 networks for all Leela Zero reviews. The playout settings have been adjusted to provide slightly stronger reviews than with the 15x128 network.
You can now download SGF’s that have the analysis information
Right now we believe the review strengths have been tuned to be about the same or slightly better than they were before the release. We are planning on giving a boost to all KataGo review strengths, but we want to see how the servers hold up to the addition of the 9x9 and 13x13 games first.
As always, let us know if you have any issues with the new system. I have fixed the things I know I broke this morning, but I’m sure there are still some issues lurking.
that’s awesome news, thank you to all you developers of this site!
Is there still an advantage/extra for the site supporters in terms of AI or does everybody get the best now?
It works the same way as before, non site supporters get the top 3 moves reviewed - although I think we do better picking those now - and we also added a quick win-rate chart for them. Site supporters get full reviews, meaning every move is reviewed.
I have to say, the AI feature just keeps getting better. Thanks for all you do to keep OGS cranking, @anoek! Out of curiosity, how long did it take you to develop the first release some time back? Did you develop this feature primarily from scratch?
It took a couple of months IIRC, though I get pulled off to fix more pressing bugs and deal with fires so actual time spent was less. Same with this release, as it was pretty much a rewrite in a lot of ways. And yeah the first one was from scratch, the second I was able to re-use a bit of the UI as a template, but there wasn’t much I left unchanged.
Love it, thanks. Congratulations. I imagine some hard work and long nights getting it all working!. I know it has only just come out and you will be iterating, but any plans to display Katago territory forecast? Can help see where the edge vanishes in handicap games…
Hey! Thanks for all of the hard work on the AI engine! Something I’d love to see is AI extended into larger boards like 21x21 and 25x25. Even if the model for these was way less tuned, it’d provide some interesting insights into the game I think. Anyway, I’d love to see that
KataGo is fairly special in it’s support for boards other than 19x19, but unfortunately I think it only goes down, not up, in supported board sizes. If that should change in the future, or another advanced engine comes out that supports larger sizes, we’ll certainly consider it - although at this time we’re not going to make any promises about other board sizes.
When I was looking at the leela zero code I believe there is now an option to train other go boards. But that does take a lot of computing power… oh well, thanks for the consideration.
Yep that’s true, though 19x19 took a metric ton of computing power, over two years by the community, 21x21 and 25x25 would take exponentially more, it would be a major undertaking.
Doing something like this, which is effectively replacing our sub-par score estimation function with a much superior kata go implementation, is something I want to do, but it’s not quite so simple to do it with low enough latency and reliability yet. It’s on the mid-long term agenda to do though I think.
Maybe I will join the KataGo project and try to extend it to larger boards. Would love to see that be a reality, both as a go player and as a programmer.
I would dislike having strong AI provide any estimation about an ongoing game (whether it be score, territory, or even just win probability) to the players of the game.
If the score estimator were to be improved with strong AI, I would hope that the players of the game would not be able to access it (during the game), or that the feature would only be present for unranked games.
On another note, I actually hope that the “ongoing game score estimator” (I only mean the one that is available to players while the game is still ongoing, since using AI as tool to help in stone status and scoring disputes is different and potentially helpful) is not improved with strong AI. Right now, the score estimator is terrible, so I don’t think too many people take it seriously and it doesn’t impact too many games. However, if players had the ability to query a strong AI to evaluate the current position, then I think that would be a form of assistance that should not be allowed.
For example, sometimes my opponent would make a blunder that I would not realize and fail to capitalize on (essentially one blunder followed by another). However, with a strong AI giving score (or even just win probability) estimates, that might tip me off that my opponent’s last move presents an opportunity that I should search for, and thus a decent score estimator could provide unfair AI assistance.
