KataGo Distributed Training and new networks

[And]

7 and 8 game 2:0 (3:5)
(TIME_B: 1825.3 TIME_W: 1867.4)
(TIME_B: 1902.4 TIME_W: 1969.4)

[lightvector]

When testing such small number of playouts, take care that numSearchThreads is set to a very small number too. I would recommend to set the number of search threads to no more than 1/20th of the number of playouts. If you have too many threads competing to perform very few playouts while the search tree is very small, many threads will be forced to search garbage variations, which could result in even worse play than having fewer threads and using fewer playouts.

[And]

this is useful information, thank you very much!
ps in these games, the number of search threads = 3

[And]

9 and 10 game 1:1
(TIME_B: 1929.8 TIME_W: 2416.8)
(TIME_B: 2099.1 TIME_W: 1584.4)

after 10 games katago b40c256-s118 50 playouts (B) - b20c256x2-s530 100 playouts (W) 4:6

[And]

katago b40c256-s118 50 playouts (W) - b20c256x2-s530 100 playouts (B) 9:1
after 20 games katago b40c256-s118 50 playouts - b20c256x2-s530 100 playouts 13:7

b20 (B) won the 7th game (TIME_B: 1791 TIME_W: 2049.1)

[And]

10 sec/move katago b40-s118 - b20-s530 15:5 (b40 won 7 Black, 8 White. KataGo - Eigen AVX2, numSearchThreads = 1)
b20 won the shortest and longest game with white:

[And]

b40 also won the shortest and longest game with white:

[And]

KataGo b60c320-s632 maxPlayouts 1, numSearchThreads 1 - CS Zero 9d H5

[And]

KataGo b60c320-s647 maxPlayouts 1, numSearchThreads 1 - CS Zero 9d

[And]

new 18 block network
https://cdn.discordapp.com/attachments/ ... r05.bin.gz
does not work on KataGo 1.11.0: "This neural net requires a newer KataGo version."
https://discord.com/channels/4170221623 ... 2637639690

[And]

New Go-playing trick defeats world-class Go AI—but loses to human amateurs

Last week, a group of AI researchers published a paper outlining a method to defeat KataGo by using adversarial techniques that take advantage of KataGo's blind spots
https://arstechnica.com/information-tec ... -amateurs/

[Harleqin]

Ugh. According to the article, this just exploits that they implemented the rules as »two passes end the game«, but KataGo by itself doesn't know this. I wonder whether the authors even understand Go.

[And]

I think the end is the most interesting:

"The research shows that AI systems that seem to perform at a human level are often doing so in a very alien way, and so can fail in ways that are surprising to humans," explains Gleave. "This result is entertaining in Go, but similar failures in safety-critical systems could be dangerous."
Imagine a self-driving car AI that encounters a wildly unlikely scenario it doesn't expect, allowing a human to trick it into performing dangerous behaviors, for example. "[This research] underscores the need for better automated testing of AI systems to find worst-case failure modes," says Gleave, "not just test average-case performance."
A half-decade after AI finally triumphed over the best human Go players, the ancient game continues its influential role in machine learning. Insights into the weaknesses of Go-playing AI, once broadly applied, may even end up saving lives.

[And]

who compiled the version for Windows with support for the new 18b network, upload please!

[PC_Screen]

Drop this katago.exe file in a working katago v1.11 OpenCL folder and have it replace the old katago.exe and it should work. The new 18b net is actually a bit faster and stronger than the old 40b net despite being "smaller" (it's a new architecture) and it's still mid-training so it should grow stronger over the following weeks/months.

https://drive.google.com/file/d/1-uWOZS ... SDRQt/view