Incognito 020823

[Eduard Nemeth]

UCI Engine INCOGNITO. This is an experimental version of CorChess with very large network (dimension 3072). The network "nn-4ffa203f3b58.nnue" was released today and I tested it. Especially with very long analyses, the evaluation in the middle game is very interesting. CorChess has a slower search than SF dev and will find more with enough time. I love it! MinimumThinkingTime, Polyglot books and self-learning in Eman style are implemented. External network. Remember: The network is very large. Therefore this engine is not suitable for very slow computers!

Download for Windows 64-Bit (AVX2, BMI2, MODERN, Source Code), Incognito 020823: Pixeldrain & Fielhorst.de --> on my homepage:
https://solistachess.jimdosite.com/solista-news/

[deeds]

...Especially with very long analyses...
...self-learning in Eman style are implemented...

What is the Interest of a learning feature for very long analysis ?
Do you mean the users will have to play 500 games/position where he did the very long analyse ?
Else the experience file will only contain score/depth for one or two moves per position.
Without count values => no experience data => it will be like an opening book with score/depth but not an experience file.

[Anton101]

Yes, it works very well for CorChess and for long time control games...I hope they don't test it against TC 1+1...The problem is that later they mix apples with oranges (I mean the Cubail test). Last night I tried it in private with TC 5+1, it's wonderful.

For openning books creation could be great.

[deeds]

...I hope they don't test it against TC 1+1...The problem is that later they mix apples with oranges (I mean the Cubail test)...

Cubail's goal never was to show the engine strength, he only wants to get a specific ranking for engines he likes or not.

[cubail]

yes
play chess room play i am

it gives an idea about chess engines, albeit simple

[Anton101]

It was clarified that it uses a larger network... That's mixing oranges with apples.

For example, this incognito network is dimension 3072...Will it also benefit to face a network of dimension 2048 on a TC of 1+1? Quite possibly not.

The test framework is for VLTC scaling, take a look: https://tests.stockfishchess.org/tests/ ... c21c0d0cee

[cubail]

15 move draw kıller book test
nnue learning has little effect

[cubail]

viewtopic.php?f=4&p=28778#p28778 look
same test tıme chance number1 chance

[Amos 4ever]

UCI Engine INCOGNITO. This is an experimental version of CorChess with very large network (dimension 3072). The network "nn-4ffa203f3b58.nnue" was released today and I tested it. Especially with very long analyses, the evaluation in the middle game is very interesting. CorChess has a slower search than SF dev and will find more with enough time. I love it! MinimumThinkingTime, Polyglot books and self-learning in Eman style are implemented. External network. Remember: The network is very large. Therefore this engine is not suitable for very slow computers!

Download for Windows 64-Bit (AVX2, BMI2, MODERN, Source Code), Incognito 020823: Pixeldrain & Fielhorst.de --> on my homepage:
https://solistachess.jimdosite.com/solista-news/

Hi Eduard,

position.cpp file
all Zobrist keys are missing..
Also remember not to implement this patch: "Remove Zobrist::noPawns"
https://github.com/official-stockfish/S ... 962b624e31

Regards

[Eduard Nemeth]

Thanks Marco! i will change it again!