Here is the SGF.
https://drive.google.com/file/d/1e6TOrQ ... sp=sharing
At Black 109, White 11-13 is a good move for KataGo. KataGo claims that this move prevents White from both ladders.
However, there is a good move on the black side that surpasses this white good move. It's black 9-11.
The bifurcation diagram is shown in the SGF, and black can take either white. However, KataGo has not been able to understand that Black 9-11 is good at White 11-13. KataGo finally notices after a few steps.
However, by tweaking the settings of KataGo, it seems that KataGo can immediately discover the good move of Black 9-11 at the time of White 11-13.
This is written in KataGo's Issues, but what kind of setting it is is kept secret.
If anyone is interested in this issue and could find a setting to get KataGo to find a good move quickly, could you please let me know?
thank you.
A brilliant move that is difficult to discover
- Joaz Banbeck
- Judan
- Posts: 5546
- Joined: Sun Dec 06, 2009 11:30 am
- Rank: 1D AGA
- GD Posts: 1512
- Kaya handle: Test
- Location: Banbeck Vale
- Has thanked: 1080 times
- Been thanked: 1434 times
Re: A brilliant move that is difficult to discover
Here is the SGF - ready to play.
Help make L19 more organized. Make an index: https://lifein19x19.com/viewtopic.php?f=14&t=5207
-
hope366
- Dies in gote
- Posts: 20
- Joined: Wed Jun 10, 2020 8:08 am
- Rank: KGS3dan
- GD Posts: 0
- Been thanked: 3 times
Re: A brilliant move that is difficult to discover
Take a look at the Go board above provided by Joaz Banbeck. In Black 109, KataGo claims White L-7 is the best move.
However, with KataGo, which is the default setting, even if you use a 60 block net, the analysis result is that black leads 4 to 5 points.
In this aspect, there is a nice move from black associated with both ladders. If KataGo can discover this move, the black lead will be +10-11 points. With the default settings, this move is very difficult for KataGo to find.
By using the configuration file prepared here, you may be able to discover the wonderful movement of black with a relatively small number of searches (about 10k).
The WRN, cpuct, FPU, etc. are set with surprisingly precise settings of 13 digits or more after the decimal point.
Please note that the memory usage has been significantly increased, which may put a heavy load on the PC.
Configuration for kata1-b60c320-s6093801984-d2892798921
https://github.com/lightvector/KataGo/f ... mmy_11.txt
Configuration for kata1-b60c320-s6119310080-d2899227642
https://github.com/lightvector/KataGo/f ... mmy_12.txt
Configuration file for kata1-b60c320-s6144619008-d2905897817
https://github.com/lightvector/KataGo/f ... mmy_13.txt
However, with KataGo, which is the default setting, even if you use a 60 block net, the analysis result is that black leads 4 to 5 points.
In this aspect, there is a nice move from black associated with both ladders. If KataGo can discover this move, the black lead will be +10-11 points. With the default settings, this move is very difficult for KataGo to find.
By using the configuration file prepared here, you may be able to discover the wonderful movement of black with a relatively small number of searches (about 10k).
The WRN, cpuct, FPU, etc. are set with surprisingly precise settings of 13 digits or more after the decimal point.
Please note that the memory usage has been significantly increased, which may put a heavy load on the PC.
Configuration for kata1-b60c320-s6093801984-d2892798921
https://github.com/lightvector/KataGo/f ... mmy_11.txt
Configuration for kata1-b60c320-s6119310080-d2899227642
https://github.com/lightvector/KataGo/f ... mmy_12.txt
Configuration file for kata1-b60c320-s6144619008-d2905897817
https://github.com/lightvector/KataGo/f ... mmy_13.txt