cheating

[flok]

Hi,

I'm not 100% sure but I might have asked this before

Let's say I built a cluster that generates billions of fens for creating a neural net. I setup and created a database for this, wrote software to talk to the chess program, created software that allows multiple computers over the internet produce the fen-strings, made the software that filters out the unrelevant fens (opening, check, etc) and dumps all of that into a large file that can be postprocessed into a weights-file by 'bullet'. Now the chess-program I want to use is stockfish as it reaches great depths in little time.
Is this cheating? Because the databse I did not write myself (mariadb), the chess-engine I did not write myself and neither the software to combine the fens into a weights-file. On the other hand I would've created all infrastructure etc etc to do all of this.

What do you think?

Please note that for Dog, I used fen-strings generated by Dog itself as that was more convenient at that point in time (full control over input and output).

[ZirconiumX]

If you intend to use Stockfish to score the FENs, I would consider this neural net laundering, which is frowned upon. At that point, why not directly use Leela data?

[flok]

If you intend to use Stockfish to score the FENs, I would consider this neural net laundering, which is frowned upon. At that point, why not directly use Leela data?

Well the net would not be compatible in this specific case. But apart from that: why not? It is not a literal copy?

[ZirconiumX]

If you intend to use Stockfish to score the FENs, I would consider this neural net laundering, which is frowned upon. At that point, why not directly use Leela data?

Well the net would not be compatible in this specific case.

...Leela training data, not the net itself.

But apart from that: why not? It is not a literal copy?

But it still inherits all the knowledge that stockfish has captured. which is the knowledge leela has captured. thus if you're going to shortcut the process by using a different engine, you might as well use leela data, which is the strongest known so far.

[smatovic]

Older thread on this:

One for Andrew Grant et al. - NNUE?
viewtopic.php?t=76386

In the Discord channels you will probably get a clearer picture what is considered original work and what not.

--
Srdja