[tools] Discrepancy between bestmove and nextmove during data generation. (not a bug)

official-stockfish / Stockfish

A free and strong UCI chess engine

https://stockfishchess.org/

GNU General Public License v3.0

11.28k stars 2.25k forks source link

[tools] Discrepancy between bestmove and nextmove during data generation. (not a bug) #3762

Open Sopel97 opened 2 years ago

Sopel97 commented 2 years ago

The move saved during data generation is the first move from PV, but the chosen move is sometimes random (how random depends on the settings). This can harm the compression of the generated data. It's not clear whether this should be touched though, as the impact might be minor. There are 2 things that should be investigated before:

how much is the compression impacted for various (popular) settings, especially multipv
how much is the training impacted, as we determine positions 'quietness' based on the stored bestmove

vondele commented 2 years ago

With the new arch, probably tools needs an update, and maybe even for the binpack 2.0 format ? Maybe you can post the link here to the design document, if it is still active?

Sopel97 commented 2 years ago

The spec draft for "binpack 2.0" can be found here: https://docs.google.com/document/d/14KuSaQbL_9Ym5IVzIDHBwGWSrGvaajc-eki8ZJ7h1Qs/edit#heading=h.wz1dp9fb5q66

It is active but no ETA. It's fairly complex to incorporate this into the current codebase in a nice way, given how many additional options there will be. We might end up having to pretty much duplicate the generation/transform code to get the most out of it. I haven't yet started implementing it.