After optimizing the parameters for one of my decks, intervals for cards that are quite mature are dramatically shortened. Even answering good on the example below gives an 11-month interval, much shorter than previous intervals, which have all ended in correct answers.
Desired retention is at 0.9.
These are the loss values before and after optimizing:
Before - Log loss: 0.3111, RMSE(bins): 6.24%
After - Log loss: 0.2863, RMSE(bins): 6.11%
Parameters: 1.5584, 1.6469, 1.7892, 7.9731, 5.0694, 1.7857, 0.3446, 0.0708, 1.6345, 0.5662, 0.9375, 2.4600, 0.0472, 0.5202, 1.2308, 0.0953, 3.9999
With these results, optimizing the parameters seems to make little sense, as cards that I remember well would be repeated too often.
I wonder if it might be caused by an irregular review history. Is it currently possible to limit the review history that FSRS uses to optimize its parameters to more recent history? This would also be convenient in case of a change in behaviour (related to the buttons, for example).