FSRS Helper - Recommended Steps

Anki Add-ons
21

Thanks to Claude, I get the stability of cards where first_rating=again and second_rating=good:

image
image1658×748 104 KB

There not exists an recommended learning step to satisfy all the three scenarios because the second learning step affects them all:

  1. If you firstly rate again and then rate good in a new card, the interval is the 2nd learning step.
  2. If you firstly rate good, the interval is the 2nd learning step.
  3. If you firstly rate hard, the interval is (1st learning step + 2nd learning step) / 2.

However, the stabilities are different among above three scenarios:

  • If we want to satisfy case one, the 2nd learning step should be 18h.
  • If we want to satisfy case two, the 2nd learning step should be removed
  • If we want to satisfy case three, the 2nd learning step should be 4h.

Here is another Steps Stats from a different collection:

image
image1648×718 105 KB

Here is the PR:

Feedback is welcome.

1 Like
22

Update:

image
image1672×736 101 KB

The second recommended learning step is based on the minimum of {S(Hard)* 2 - S(Again), S(Good), S(Again Then Good)} now.

2 Likes
23

Isn’t short-term memory considered to be about 20 seconds? Pretty sure these relearning steps we’re talking about are all considered long-term memory. Why are we thinking FSRS 5 can’t handle these?

24

Yeah, I think we’re talking about STM in a different sense (not in the academic one).

25

Why would we not be talking about it in the academic sense? We’re modeling memory in the academic sense.

26

Then you’ll have to invent a new term or be wordy in your description, “uh, well, our DSR model doesn’t work well for intervals that are short. we can do fine with a day or two but anything less than a day gets a bit fuzzy”

27

Does it not work well? We have data on that? Maybe I missed that part.

28

The forgetting curves behave weird. In @DerIshmaelite’s case he even remembers more over time (I have no idea why). So the curve goes up at some point.

29

Reading through this, I’m getting the impression that because FSRS isn’t designed to handle short-term memory, that fact per se is why we’re assuming FSRS 5 isn’t going to handle relearning steps. Because short-term memory functions differently. But if relearning steps are actually based on our long-term memory, in the academic sense, that assumption shouldn’t hold.

If there’s actual data that it doesn’t model relearning steps well, then this all makes sense.

30

Personally, I’m liking the way it’s scheduling cards when I answer “Again”. Hope that doesn’t go away.

31

There are too many theories about memory and I don’t think we should take all ideas from the experts too seriously.

32

You don’t think we should take seriously the understanding of what constitutes long-term/short-term memory according to the experts?

33

Why should we take the assumption that FSRS can’t schedule relearning steps well seriously? Is there any data for that?

34

I don’t think we should take the categorizations very seriously.

35

The categorizations seem like the entire reason we’re making the assumption FSRS can’t model these, because it’s not designed to model short-term memory.

36

I don’t get the sense that these conversations have hewed strictly to the 20-second academic rule as a working definition for FSRS, FWIW.

LM Sherlock has stated that FSRS5 is not meant to model short-term intervals.
I’m sure he can weigh in as to where he’s drawing that line, though it seems to be, to some degree, roughly divided along the 1d boundary line owing to Anki’s treatment of same-day versus inter-day intervals (from what I gather) and the relative difficulties in modeling either.

The consensus seems to be that modeling the forgetting curves over shorter timelines is more difficult, which is the fundamental issue at hand.

image
image785×143 12.4 KB

37

I have. But I haven’t published it. The raw data is stored here: srs-benchmark/result/FSRS-4.5-secs.jsonl at main · open-spaced-repetition/srs-benchmark

FSRS-5 is not designed to model short-term memory because the short-term reviews aren’t used to evaluate FSRS-5.

38

So the issue is with the way that FSRS reads the data on same-day reviews, yeah? This has nothing to do with whether or not the learning curve models these reviews well? I remember reading that FSRS isn’t, as of now, able to read the actual intervals on same-day reviews, it treats them all the same.

39

Based on my preliminary experimental results, if FSRS were to predict short-term memory in the same way it predicts long-term memory, the overall error would double. And it will become too conservative to predict the long-term memory.

1 Like
40

Do you have thoughts on what I mentioned earlier, short-term memory being just 20 seconds or so, and these relearning intervals are all long-term?

I’m also wondering if the error being so much larger has something to do with the history of almost everybody’s relearning steps being manually selected. There is no data on FSRS scheduling the relearning steps, so the data is going to be skewing the errors.