Can you train an AI on Stockfish evaluations

[modern vector]

give positions and the stockfish evaluation on first ten good moves, and expect to build an AI that makes good moves, without checking for legal moves etc. What do you think?

[tropic phoenix]

That sounds similar to what dkappe was doing some years back.

Depending on what you aim for, you'd still want to include this into a tree search.

You could of course also skip all of that and don't do any search at all, but that would be considerably weaker.

[fallen crystal]

Why don't you want to check for legal moves? Is the plan to only use a NN with no additional code?

In that case, I think you could simply use a Leela net, afaik it has somewhere around 2300 elo when part of a 1 node search

[modern vector]

I don't know, just looking for ideas how to integrate stockfish into training an interesting AI, but that's what NNUE is all about i guess.

[pseudo wolf]

You're not the first to have that sort of idea https://www.chess.com/forum/view/general/chess-novice-studies-a-bit-for-30-days-and-surprise-can-t-beat-magnus

I'd suggest finding a more interesting problem, like an AI coach to teach players how to play better.

[small meteor]

PGN to NLP.