CGLemon
CGLemon
Hello, Yuki Kobayashi! I am Hung-Zhe Lin. I raise the new issue here instead of email because it is more easy to write Latex here. ### On the Paper After...
Yuki Kobayashi: According to your paper, I try to add the Progressive Widening in the MCTS. But I find the PV string is too long. Here is it. It is...
This is the training result in my project. The ```fixup``` is Fixup Initialization (with gradient clipping). The ```renorm``` is Batch Renormalization. The two network both use the same training data...
In Leela Zero, the team implemented LCB by computing quantile of student's T pdf and Welford's online algorithm. It is easy to understand. But LCB algorithm is complex in KataGo....
KataGo computes the pass-alive string in order to force to finish the game if someone refusing pass. https://github.com/lightvector/KataGo/blob/6e41a3693f47eaaba6afa55ba23418b76a0675c8/cpp/game/board.cpp#L1808-L1814 Seem the pass-alive also help to remove the dead strings. But I...
There is the analysis winrate or score lead value on the graph. I think that it is more easy to get the best winrate if there are too many candidate...
Ogatak is a graceful analysis tool but there are two drawbacks for playing. One it that we can not toggle between analysis and playing modes quickly. Another is that we...
Add support for distinguishing the next move when there are 2+ moves. This feature is in the TODO list #15. Black Effect  White Effect , the best ```C_SCALE``` value is ```0.1``` on 9x9 board. Why did you use ```1```. Did you tune it by your personal experiment?