@Danika_Dakika Please merge my topic with this one so this does not get confusing.
Let’s hope Obezags’ work has a breakthrough.
However, I don’t understand on what basis is a preset defined here 
That is the whole purpose of making this topic in the first place: we don’t what “secret-golden-preset card distribution” there is to ensure minimal RMSE overall increase.
It’s just based on the user’s presets. So if the user has 2 presets, FSRS will be optimized on 2 presets. If the user has 200 presets, FSRS will be optimized on 200 presets. Each with their own parameters.
Wasn’t it always like that, like forever? Or was FSRS benchmarking based on one general preset?
I don’t see what is the use if the number of presets is not fixed to anything (like an optimum number of presets ).
All benchmarking results up until now were based on using the same parameters for the entire collection.
The use is knowing that “naively” optimizing FSRS on every preset is not a bad idea…on average, at least.
And how many presets could you possibly “naively” make without it damaging RMSE
No clue
I’m sure there’s no need. You’ve linked them by reference, and there’s no reason to noise-up someone else’s topic.
Another update: while per-preset optimization may be very slightly beneficial, per-deck optimization is definitely not.
FSRS-5 per deck performs worse than FSRS-4.5. This supports the idea that users are already decent at choosing which preset to assign to which decks.
Btw, FSRS-5 presets is better than FSRS-5 for 57.1% of users.
