Can Stockfish evaluate the position in which it has not yet been trained?

[ChessOverflow]

Hello! I have recently learned about neural network engine, but I have little knowledge of how it works. I want to know if Stockfish can correctly evaluate positions that it has not trained yet?

[cj5716]

you cannot train a neural network on all positions possible to be reached on a chess board. However, a neural network "learns" from the data it was trained on, so that it can give an accurate evaluation even on positions it was not trained on.
see https://github.com/official-stockfish/Stockfish/wiki/Advanced-topics#classical-versus-nnue-evaluation and https://github.com/glinscott/nnue-pytorch/blob/master/docs/nnue.md

[dav1312]

This is called "generalization". It refers to the ability of a model to apply concepts learned during training to new data that was not seen during training.

[ChessOverflow]

Do you mean by generalization the same as reinforcement learning?

[dav1312]

No

[ChessOverflow]

How can I download all positions that have been trained so far?

How can I detect if a position has been trained or not?

[cj5716]

Training data here: https://robotmoon.com/nnue-training-data/