Summary:

* New run learns twice as fast as the old LZ130 run and now reaches ELFv2 level at visit parity in about 19 days, and using no more than 27 GPUs for most of the run (19 for early parts, 28 briefly) instead of hundreds or thousands as in Leela Zero/ELF/AlphaZero. Self-play games and training data all released: https://d3dndmfyhecmj0.cloudfront.net/g104/index.html

* Major steps forward on OpenCL and Windows support. More work still to be done, I'm working on it.

* The OpenCL/Windows issue aside, should work fine with Lizzie and other tools at this point (and if not, I should be able to make any fixes quickly and easily if it's just a matter of outputting the data the right way for a tool to read). Capable of also predicting final score and territory heatmap, as soon as any tool wants to add support for how that tool wants to display it.

* Still have many research ideas for how to continue to improve it. I'll be continuing to test these these ideas over the next months, and if any of these ideas test out well, we should be able to do an even better third run.

I'm happy to answer any questions or whatever. Yay.

@lightvector
Wanted to congratulate you and the entire KataGo team with the 1.1 release.
I am convinced that it will have a big impact on the AI study tools.

Installed it this evening on an ubuntu test server. Looks promising.
Will play around with it in the coming days.

PS: are you also "@lightvector" on twitter ?

_________________
Enjoy LeeLaZero and KataGo from your webbrowser, without installing anything !
https://www.zbaduk.com

Thanks! Seems like the main big remaining obstacles to usability are Windows support and CUDA dependence. Hope to address both of those over the next few weeks.

No, I am not on Twitter at all.

Very exciting … any chance we can get this for macOS?

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

Or an OS-indep browser version like deepleela or zbaduk ( greedy )

… or http://leela-one-playout.herokuapp.com/ (had to add it to the list)

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

I have a friend who I think has managed to compile it for mac already, so I think the answer is already yes? Let me know if you run into problems, happy to troubleshoot. If you're looking for pre-compiled executables, I'll try to figure that out once the OpenCL implementation is fast enough for real use, since figuring out how to link with CUDA and deal with CUDNN is a headache, even sometimes right on the same OS I do all my development on (linux).



Well, I hope whatever dev or devs like to develop sites like that in their spare time find KataGo worth picking up.

Oh wow, sounds good … but meanwhile I’m elderly and not so adventurous as I once was … so, I’ll wait for click-ready executables

Why not talk to Andreas M. Hauenstein about it? His mail address is on the “About” page.

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

I've pushed some code to the q5go github to support kata-analyze and save/display the score data it reports (it's added to the comment in batch analysis, and the evaluation graph now has a context menu item to switch between score and win rate).

Compared to ELFv2, Katago's handicap game style is really the best. It can play 5 handicap games ( I haven't test more handicaps yet) without any bad moves (remember ELF's ladder moves?).

Here is a 4 stone handicap game with Leela 0.11 with 20 second per move time setting. With the same time setting Leela Zero 40b weights have only small chance to win in 3 handicap games. The obvious style of Katago is that it likes to cut and deal with multiple groups properly. It even outsmarts Leela 0.11 in ladder reading.

Here is an example of a fascinating game between LZ and KataGo I played out tonight in Sabaki while working out. Hey, gotta watch something, right? It's like baduk TV!

The hardware was:
AMD Ryzen 2600x
NVIDIA RTX 2060

The settings were as follows:

KataGo 1.1


Leela Zero #226


I made the first four moves myself, because I didn't want to watch a bunch of 3-3 invasions this time. The game features an early ko and a complicated fight in the center that decides the entire game. Neither engine was sure who was winning until the very end.

I find high level bot v bot games both confusing and amazing. The midgame fights that break out can become so complicated.

As far as I understand, KataGo needs CUDA/CuDNN, which requires NVIDIA GPUs, so there might be no way to run the current version of KataGo on Apple hardware (macOS or any other OS), since Macs use Radeon GPUs.

In this discussion there has been mention of an OpenCL version, so I assume that doesn't require CUDA etc., right?

FWIW, my 2014 Macbook Pro has an NVIDIA GPU:

"Chipset Model: NVIDIA GeForce GT 750M"

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

That is correct. It will still need an OpenCL device, normally a gpu but also supported by some cpus (although that will be relatively slow).

Now this is a strange game between Katago(B) and LZ230(W). Near the end of the game white played F8 so that there was a seki in ko situation.



Both sides think they are losing. Maybe Katago thought the ko was too big? Maybe LZ thought it was a bad ko to start? So at least one of the bots was wrong and finally LZ230 resigned. I hope someone can explain this because I'm not a strong go player.

splee99, it cannot be a seki because White H8 string is dead. White's only option is to start the ko by playing A5, and Black takes first. By the very end I don't see any White ko threats, so Black will win the ko.

For Black maybe doing something earlier to resolve the situation is safer -- either taking the ko (and filling it to make a temporary seki), or filling liberties of H8. But maybe Black calculated White starting the ko isn't dangerous since Black will take first.

This is an approach ko ( https://senseis.xmp.net/?ApproachKo ). White cannot win this ko. Earlier in the game it may have been better for White to convert this to a direct ko and gain something in exchange for losing the ko. FWIW, I am unaware of any bot ever playing this strategy. I am not sure that current bots could learn it, but in any case they might be able to read the ko fight out and play it.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.