For me the FSRS (FSRS 6 on latest AnkiDroid 2.22.3) difficulty for difficult cards rises in the learning phase as it should (often even to e.g. 98%), but when entering the review phase this is a problem, b/c the review intervals become way to short. Even when I click easy, the difficulty is only reduced slightly (e.g. 97%). Once these difficult cards make it into my memory for a day, they become much easier and the review intervals should be much longer than they are right now. Thank you very much for your attention and any comments!
Big thanks to all Anki contributers - Helge
AnkiDroid Version = 2.22.3 (4f2ec681da36b8307a865f05192498b8425adf0f)
Backend Version = 0.1.60-anki25.07.5 (25.07.5 7172b2d26684c7ef9d10e249bd43dc5bf73ae00c)
FSRS = 4.1.1 (Enabled: true)
It would help debug if you pasted your FSRS parameters. You can graph how your intervals change yourself using this tool if you paste your parameters in:
The tool is amazing! With the tool I can hand-optimize my parameters. I had hoped that the optimization could do that for me :(. But probably it cannot, b/c with high difficulty cards it never gives me the chance to try to remember after a long interval, b/c it only gives me short intervals even on easy.
Maybe the problem is that my cards vary very much in their difficulty. Trying to sort them into different decks doesn’t feel right b/c there are no real categories, it is a continuum.
By the way, b/c you asked, here are my (unsatisfying) params from optimization on 1500 reviews: 0.0150, 0.0599, 0.1566, 0.3891, 6.5344, 0.6964, 3.1485, 0.0010, 1.6894, 0.2517, 0.6083, 1.3899, 0.0647, 0.1413, 1.5737, 0.5031, 1.6950, 0.5631, 0.1458, 0.1205, 0.1230
Please don’t manually modify your parameters. The optimizer does better than a human can at fitting the parameters to your review history. Instead, lower your desired retention if you find that your intervals are too short (I feel I should have lead with that :P).
Thank you for your advice! Changing my desired retention does not work for me, because the intervals are only too short for cards with a high difficulty. For cards with a low difficulty they are okay or even too high.
The FSRS curve fitter does an infinitely better job than any human could ever dream of doing. Do not second-guess it. Click “Optimize preset” every now and then. Set your DR.
If it has data for a given region in the parameter space (which… it seems it does)… it is going to give you the correct parameters.
Second-guessing it is useless. It can only hurt you.
I found a solution for me I think: I reset the progress of cards for which AnkiDroid offers me a way too short interval on the easy button and then set an appropriate due date. Now the difficulty has been reset and when the due date comes the interval on the easy button will be normal and fine.
At this point an auto response should be made for those questions, which are legitimate, but often take the assumption that D should be able to revert back to a "lower” value with time.
Thing is, D behaviour is dependent on params that are just going to be optimized just like any other params. And for many of us, doing a few mistakes, leading to a huge D=98%, which will never go back to a 40%, is … what’s the most accurate way to represent how our cards behave.
Now, why do they behave like this ? Because essentially, you probably have 2 groups of cards : The cards that you never fail, and the cards you will keep on lapsing, even with shorter intervals. This might be due to bad “knowledge stability” (see it as “guessing answer” and getting it right X% of the time) which lead to non increasing stability over reps. FSRS will just adapt to it and flag those with high D and lock them there.
By resetting those cards to allow them to have low D again if you feel you won’t fail them anymore, you basically reinforce that behaviour : either a card will fail for the foreseeable future, or it won’t ever.
IMO the best approach is just to do nothing. Don’t see the clusters of D (around 40%, 60%, 80% and the 97%) as a distribution but as a clustering that is done based on D. Then, inside each clusters, you indeed have a normal distribution. If you "zoom” on D distribution (some addons allow you to see it like Search Stats Extended), you’ll notice how 96% D vs 98% D can make a big diff on how stability will increase. Of course, never like D=40%, but it’s not like they are completely locked either way.
Bottom line : Those params don’t have those values because they are bugged, they have those values because that’s what describe the better how you actually perform with cards.
I hope FSRS7 will solve my problem, because expertium wrote in another thread that it will handle same day reviews properly. Add non-linear grade-delta-difficulty
Looking at the formulas of FSRS6, I don’t see how it could adapt to the case where a user finds the card very difficult on the first day of learning, but can then mostly recall it from long-term memory much easier.
I now have a second case where the problem is occurring in exactly the same way. My daughter is learning English vocabulary with FSRS and she’s running into the exact same problem, namely that she needs a lot of repetitions for the initial learning, which drives the difficulty to a very high value, but after that, as soon as she has known the vocabulary for a day or two, the intervals are way too short, because now it’s in her long-term memory and she can remember it for much, much longer.
But I know now that this is a known problem and @Expertium it’s working on it. He says same-day learning has only been implemented with a rough heuristic so far (in FSRS 6).
