Worse results after tuning

[Rentib]

What do you do if after tuning the engine performs much worse?
Do you generate more data or completely replace the existing one with another set?
Or maybe you remove some parameters from evaluation function that seem to have really stupid values after tuning?

I tuned my engine on ~4million quiet positions from selfplay and it lost around 160 elo after runing adam optimization for a while, so I am a bit confused as to what to do...

[JacquesRW]

What do you do if after tuning the engine performs much worse?
Do you generate more data or completely replace the existing one with another set?
Or maybe you remove some parameters from evaluation function that seem to have really stupid values after tuning?

I tuned my engine on ~4million quiet positions from selfplay and it lost around 160 elo after runing adam optimization for a while, so I am a bit confused as to what to do...

You look carefully through your datagen and tuning code for bugs. You don't just drop 160 elo out of nowhere with tuning.
Make sure you're filtering your selfgen data appropriately as well:
- positions with mate scores
- positions in check
- noisy positions (there's a couple ways to classify "noisy", e.g. best move is a capture or qsearch differs from static eval by more than some threshold).

[Rentib]

To obtain data I run ~64'000 games with some starting positions with tc 1+0.08.
Then I remove 16 last positions from each game (there is no way my engine will do a 16 depth search).
After that I remove positions where static evaluation is different than score returned by quiescence search.

I agree that I might have a bug somewhere, because this loss is just too big.

[op12no2]

just a thought; you could try training against known good datasets as a test. for example quiet-labeled.epd and lichess-big3-resolved.epd. I think that's about 8M positions with WDL (but no scores).