TalkChess.com

dkl wrote: ↑Thu Jan 07, 2021 11:38 am I was very interested in how the NNUE is actually designed an implemented.

Yu Nasu originally introduced it in his 2018 paper "Efficiently Updatable Neural-Network-basedEvaluation Functions for Computer Shogi" (available here https://www.apply.computer-shogi.org/wc ... D/nnue.pdf ), but it's unfortunately only available in Japanese. Only the abstract is in English.

I thus created a rough translation of this paper to German and English. The PDFs are available at my github repo here
https://github.com/asdfjkl/nnue

The German translation is quite literal word-by-word, for English I took more freedom. Of course, if someone spots translation errors, I'll try to fix these over time. Some sections that might seem confusing are however already in the original paper difficult to read, as they assume some knowledge about certain computer Shogi aspects (like the previously commonly used Shogi evaluation function, or the KP encoding).

Still though it might be easier to understand this from the paper than from looking at code. It might also make sense to link these in the chessprogramming wiki...?!

dkl wrote: ↑Thu Jan 07, 2021 11:38 am I was very interested in how the NNUE is actually designed an implemented.

Yu Nasu originally introduced it in his 2018 paper "Efficiently Updatable Neural-Network-basedEvaluation Functions for Computer Shogi" (available here https://www.apply.computer-shogi.org/wc ... D/nnue.pdf ), but it's unfortunately only available in Japanese. Only the abstract is in English.

I thus created a rough translation of this paper to German and English. The PDFs are available at my github repo here
https://github.com/asdfjkl/nnue

The German translation is quite literal word-by-word, for English I took more freedom. Of course, if someone spots translation errors, I'll try to fix these over time. Some sections that might seem confusing are however already in the original paper difficult to read, as they assume some knowledge about certain computer Shogi aspects (like the previously commonly used Shogi evaluation function, or the KP encoding).

Still though it might be easier to understand this from the paper than from looking at code. It might also make sense to link these in the chessprogramming wiki...?!

dkl wrote: ↑Thu Jan 07, 2021 11:38 am I was very interested in how the NNUE is actually designed an implemented.

Yu Nasu originally introduced it in his 2018 paper "Efficiently Updatable Neural-Network-basedEvaluation Functions for Computer Shogi" (available here https://www.apply.computer-shogi.org/wc ... D/nnue.pdf ), but it's unfortunately only available in Japanese. Only the abstract is in English.

I thus created a rough translation of this paper to German and English. The PDFs are available at my github repo here
https://github.com/asdfjkl/nnue

The German translation is quite literal word-by-word, for English I took more freedom. Of course, if someone spots translation errors, I'll try to fix these over time. Some sections that might seem confusing are however already in the original paper difficult to read, as they assume some knowledge about certain computer Shogi aspects (like the previously commonly used Shogi evaluation function, or the KP encoding).

Still though it might be easier to understand this from the paper than from looking at code. It might also make sense to link these in the chessprogramming wiki...?!