Lczero stalled in progress?

[corres]

I think the developers of Leela Chess Zero should learn methodology from the developers of Stockfish.
The success of Stockfish originates mainly from its systematical development.

[Michel]

I think the developers of Leela Chess Zero should learn methodology from the developers of Stockfish.
The success of Stockfish originates mainly from its systematical development.

That is easier said than done. The feedback loop of lczero is much longer than that of SF since it passes through the training window (500000 games I believe). Some patches like FPU reduction may appear to gain elo but one does not know how they affect the training of further nets (FPU reduction is an interesting case since it interferes with the normal functioning of the policy head). Personally I would prefer to keep the training version of lczero as pure as possible.

BTW SF development was not created out of thin air. The principles on which it is based (additivity of small elo patches, short time control testing, random opening books, ...) were all controversial once and it took a long time before people became truly confident in them. We still have to learn to what extent those principles apply to the radically different NN/MCTS engines.

[corres]

[quote=Michel post_id=762048 time=1526203562 user_id=4139]

[quote=corres post_id=762042 time=1526199493 user_id=8071]
I think the developers of Leela Chess Zero should learn methodology from the developers of Stockfish.
The success of Stockfish originates mainly from its systematical development.
[/quote]

That is easier said than done. The feedback loop of lczero is much longer than that of SF since it passes through the training window (500000 games I believe). Some patches like FPU reduction may appear to gain elo but one does not know how they affect the training of further nets (FPU reduction is an interesting case since it interferes with the normal functioning of the policy head). Personally I would prefer to keep the training version of lczero as pure as possible.
BTW SF development was not created out of thin air. The principles on which it is based (additivity of small elo patches, short time control testing, random opening books, ...) were all controversial once and it took a long time before people became truly confident in them. We still have to learn to what extent those principles apply to the radically different NN/MCTS engines.
[/quote]

But just causes you lined up the developers of LC0 would be more cautious when they make modifications.
Even if you know every each modifications they have positive effect you can not know that if you applies all modifications simultaneously the result will be good also.

[Robert Flesher]

It would appear LC0 is once again seeing progress. Good stuff!