Expertium

As for the idea I had in mind, I benchmarked it and it didn't work. https://github.com/open-spaced-repetition/fsrs4anki/issues/461#issuecomment-1859140836

> * Calculation of difficulty after review should depend upon the retrievability, just like in SuperMemo algorithms. While that would be great, all my attempts to integrate R into D...

@L-M-Sherlock please make a version of the optimizer with visible code so I can test this idea? I need it for other testing as well, I can't test much if...

I tested it using this file (I made a copy): https://colab.research.google.com/github/open-spaced-repetition/fsrs4anki/blob/main/archive/candidate/fsrs4anki_optimizer_beta.ipynb I added 2 more parameters, like this: `torch.exp(-(retrievability - self.w[15]) * (grade - self.w[16]))` Also, I had to modify...

Initial D doesn't matter as much as initial S. I would really like to see a hybrid model, with all FSRS formulas, except that D is given by an LSTM....

I don't know how to test it the "proper" way, so I tested it my way again: I made another copy of https://colab.research.google.com/github/open-spaced-repetition/fsrs4anki/blob/main/archive/candidate/fsrs4anki_optimizer_beta.ipynb and implemented the code there. It performed...

LSTMs are specifically designed for time series. Input: interval lengths and grades, output: a number between 1 and 10, which you then use in formulas for S.

Yes. The LSTM should replace `new_d`.

@L-M-Sherlock just a reminder to implement my idea with LSTM. Input: grades and interval lengths, output: a number between 1 and 10. It should replace the formula for `new_d`. Why...

I still think that we should try what I suggested here: https://github.com/open-spaced-repetition/fsrs4anki/issues/352#issuecomment-1678869322, https://github.com/open-spaced-repetition/fsrs4anki/issues/352#issuecomment-1678889261 If an LSTM cannot generate a value between 1 and 10 that, once plugged into existing formulas,...