CGLemon

Results 40 comments of CGLemon

不好意思,哪裡有提及 KataGo 只使用十幾萬盤遊戲呢?

training games 代表用這個權重產生多少盤訓練資料,rating games 跑多少用以測定強度的對局,所以並不是前資料集大小。現在 KataGo 至少使用最新的數百萬盤作為訓練資料,詳細數字我就不確定了。 這個版本主要改進為支援日本規則,因為 UEC cup 是使用日本規則,看現在時間接近了,趕快丟上去練,餘下的是一些關於強化學習的小改進錯誤修正。 另外一個比較重要的改進為,嘗試支援強度調整和多變棋風,看起來兩者結合能提昇對戰體驗,但當前 v0.7.0 只有開放強度調整選項而已,預計 v0.7.1 後才會整合棋風選項。

The engine can't play the move in the pass-alive and pass-dead area. So the engine can't remove all dead stones sometimes. I will add this feature into TODO llist.

Add support for ```---capture-all-dead``` in the branch [remove-dead-stones](https://github.com/CGLemon/Sayuri/tree/remove-dead-stones). Look like it works well. The version is v0.5. Old support for old netowk, such as UEC14 and CGF 2023 network.

Thanks for testing! Look like I missed some conditions. Already fixed the capture bug in the same branch. But this option is unsafe in some positions. Did Aya implement the...

Fix capture-all-dead mode bug and push the code to the master branch. Now can capture two point eye group.

Thank! A little of off-topic, The Elo result shows the ```sayuri811_2035_p1k``` almost achieves ```LZ_05db_ELFv2_p800```. But in my local testing, Sayuri is significantly weaker than LZ174 which is also significantly weaker...

That is weird. I got this github [side](https://github.com/mstorsjo/llvm-mingw/releases). Where is your WinGW from? I want to test your GCC version on my computer. Thanks!

Thanks! That was mistake. I will update it soon.

Sorry, I forget to update the command. Please use following GCC command instead of the old one. ``` g++ -std=c++14 -ffast-math -I . -lpthread *.cc utils/*.cc summary/*.cc game/*.cc mcts/*.cc neural/*.cc...