Note : Fun tesuji from Igo Hatsuyoron 120 trained katago

[alain-bkr]

trafficstars

Not a bug, just for the record.

I setup a bot on kgs, with the NN from Igo Hatsuyoron study , just to see how it plays and if it had learned some specific things. (i was thinking of nakade)

Currently after one day it is near 7~8d on kgs , (lose at 2 handi against strong KataBot, win with 4H)

While watching a game at 9 handi against a 4k (who won) , i recognized one move from Igo Hatsuyoron 120 : the Guzumi :-)

A screen shot.

• on the left, IgoH120 study at the precise moment of the tesuji (which is usually not so good)
• on the right, the (lost) game, with the same move Guzumi_lol

The game was really lost, the studious bot resigned a bit later.

Conclusions :

• It is still very strong and read very well complicated fights (i am 1k)
• The NN learned something from IgoH120 (and made me laugh)

Thanks a lot for Katago.

Références :
By the author of Katago : a short history of this specific training and efforts to tackle this problem. https://blog.janestreet.com/deep-learning-the-hardest-go-problem-in-the-world/

Recent (2022) evolution and current supposed best analyse. https://igohatsuyoron120.de/2015/2022MainLine.htm

g103-b6c96-s103408384-d26419149.txt.gz , with 10s on OpenCL RX6800XT, no pondering.