allowResignation affects humanSL strength?? about katago HOT 4 CLOSED

I've tried more variations.

(OK = suitable for 9k, NG = too strong for 9k, in my impression)

• Remove kata-raw-human-nn: OK
• Insert name command just before kata-search_analyze_cancellable: OK
• Use genmove instead of kata-search_analyze_cancellable: NG
• Change the backend from OpenCL to Eigen: OK? (I may need to adjust time_settings further for a fair comparison.)

Here are the policy spectra for games that I found easy (*1 = rank_20k) and hard (*1 = rank_9d), but I haven't been able to reproduced the issue without my GUI...

from katago.

I'll close this issue since things are unclear. Sorry for bothering you.

from katago.

@kaorahi If there's a major bug in KataGo here with silently using the wrong profile despite having set another one, I'd like to fix it. In these cases you should be able to see from KataGo's own log file (with logAllGTPCommunication=true) what sequence of commands it perceived, to verify that it received all the commands you expected your GUI to send, in the expected order. Additionally, you can try using ranks that are much further separated than 1d and 9k, such as using 9d and 20k, to get a far sharper distinction so that your testing is less unclear.

Do you think there's a bug here?

from katago.

No, I'm starting to doubt the issue itself now.

Thank you for caring.

At a glance, KataGo's GTP log looks as expected. I'm beginning to think that the issue of being "too strong" might just be a coincidence or a preconception. I wasn't able to reproduce the issue in matches between the two 9k KataGo engines with the GUI settings that I considered "OK" and "NG". They are running on different machines to avoid any mutual interference.

from katago.