Xiangqi endgame play weaknesses

[shuangqiao999]

It would be nice if it could be improved to load the endgame library file. Thank you for letting everyone use the engine. Thank you

[ianfab]

Could you please provide example positions/games and illustrate the misevaluations that occur there?

[shuangqiao999]

Hello! The xiangqi engine has repeatedly captured the situation, which is in violation of the xiangqi rules will be judged negative interface

[ianfab]

Are you talking about chasing rules, like #55? Either way, examples would help to understand what you mean.

[cloudfish]

Could you please provide example positions/games and illustrate the misevaluations that occur there?

The best way to gather examples would be testing Fairy-SF against stronger engines. I plan to post a tutorial about a stronger Chinese engine on discussions or discord so that anyone interested may find some other examples.

A readily available example would be the game in #405. The endgame of this game is 车马兵有仕相必胜车士象全(KRHPA/E book win KRAAEE) in Xiangqi tutorials (see a similar example in game 261 here).

FEN after turn 61 (last capture) is 4kae2/4a4/4e4/9/PR7/9/3H4P/9/9/2E1K1r2 w 0 1 . Fairy-SF NNUE evaluates it to +0.42 at depth 50. TianTian Xiangqi evaluates it to +7.97 at depth 25.

FEN after turn 90 is 2e1ka3/4a4/4e4/7H1/P2R3P1/2E6/5r3/9/3K5/9 w 0 1 (ply information removed to eliminate influence of 50-move rule). Fairy-SF NNUE evaluates it to +2.81 at depth 36. TianTian Xiangqi evaluates it to +11.41 at depth 23.

Of course a dedicated Xiangqi engine may use hard-coded endgame knowledge here, and I'm not familiar with concepts of NNUE, but could this indicate that the training of NNUE is affected by rule issues like #405 or #55?

[shuangqiao999]

Now the best Chinese chess engine is the whirlwind engine (www.ccyclone.com) using neural network algorithm, very strong is the commercial engine unfortunately too high fees

[shuangqiao999]

Hello, can you add a configuration file for the engine? Because the existing Chinese chess interface does not have the option of whether to enable nnue or not, if you do not enable nnue and just use the classical algorithm, you cannot use the maximum power of the engine. The existing Chinese chess world also has problems with setting the hash value to 1024MB, which actually only takes up over 500 MB of memory I really hope you can publish an engine profile and thank you so much for taking the time to read my message.

[ianfab]

NNUE usage is enabled by default, so there should not be a problem if the GUI does not allow to change it. I do not fully understand the issue with the hash size, but if there is a bug or limitation in the GUI that makes it impossible to set larger hash sizes, it should be fixed there and not in the engine. If the issue is related to https://github.com/ianfab/Fairy-Stockfish/issues/412, then it should be solved with the next release.

[shuangqiao999]

Hello there is no training xiangqi nnue file tool?

[ianfab]

The NNUE training code is private.

[shuangqiao999]

Thank you I understand I thought training tools are also public please understand my ignorance 😀

[15890944033]

Hello

[15890944033]

I want to cooperate with you to develop chess

[15890944033]

I have excellent Chinese chess source code and won the fourth place in the Chinese chess competition

[15890944033]

My QQ number is 2237556025

[15890944033]

I hope you can contact me

[ianfab]

@15890944033 Hi. I do not have QQ, but feel free to open a discussion or join the Fairy-Stockfish discord to discuss more.

[hotfics]

@ianfab for example , FEN : 2bakab2/1c2n4/4r4/C6Np/9/P1P5P/6P2/4B4/4A3n/1c2KAB2 w - - 0 2 white can win the game,but fairy don't play well in this position, the key is to lock the black horse in the south-east

[hotfics]

@ianfab 2cak1b2/4ar3/1R7/9/9/9/9/9/4A4/2BAK4 w - - 0 29

white rook permanent capture black's cannon ,which is ok in chess,but not allowed in xiangqi