FSRS Button strategy for new cards: good or again?

Anki FSRS
1

Hi everyone,

I wanted to share my approach to using FSRS buttons and get your thoughts on whether it’s “correct” or if I might be undermining the algorithm’s effectiveness.

I’m using the latest Anki with FSRS, getting a 4.07% RMSE on my main Japanese language deck (radicals, kanji, vocabulary) and a 3.78% RMSE on my geography deck. I have separate FSRS parameters for different deck types.

For review cards I follow the standard approach of using again if I don’t recall the card and the other buttons otherwise (mainly good).

For new cards in my language learning deck I click “easy” if I already knew the back before reviewing the card; “good” even if I’m not too familiar, I’ll study it and click it once I think I’ll remember it in 10 minutes; never click “again” because I’d rather study it now, click “good,” and do 2 reviews total (initial + 10 min) rather than 3 reviews (initial + 1 min + 10 min).

For geography and similar content I use “again” more frequently because these types of associations are more difficult to remember without mnemonics but I still apply the same principle of clicking “good” if I think I’ll remember the card in 10 minutes.

I do this because I find it inefficient to use “again” and see the card in 1 minute if I’ve just studied it and will still remember it in 10 minutes. I’d rather do 2 reviews total for new cards than 3 with what feels like a redundant step in between.

I’m achieving my target retention rates, and the FSRS parameters seem to be optimizing well (4.07% RMSE for my Japanese deck).

I’ve seen some threads (on reddit) advising against using different button strategies for new vs. review cards. However, I’m being consistent within each stage (new cards vs. review cards), using separate FSRS parameters for different deck types and achieving good retention results.

Is this approach actually problematic? Or is it reasonable to have different (albeit systematic) criteria for new vs. review cards?
Or should I just change my learning steps from 1 minute and 10 minutes to simply 10 minutes?

My understanding is that FSRS treats new cards and review cards as different stages in the learning process anyway, so having different button criteria for each stage should be fine as long as I’m consistent within each stage.

I’ve already accumulated over 10,000 reviews with this approach in my language learning deck. If I were to change my button usage strategy now (or my learning steps), wouldn’t that potentially disrupt the algorithm’s efficiency even further?

Would love to hear from more experienced FSRS users on this.
Sorry for the wall of text :')

Thanks!

2

For anyone starting out, this seems like the most reasonable solution. You don’t need the 1m step for anything, since you’re using the buttons now to effect what it would be like if you had a single 10m step. It’s also easier not to have different grading policies for different types of cards.

For you – if this is working and you want to keep doing it – go ahead!

3

Thank you for the reply!

I tested changing the steps to “10m” only, but the behaviour is different.

When I click “again” on a card, it reappears after 10 minutes as expected. However, after graduating, its next interval is only 1 day rather than the current 5-6 days. My only learning step is “10m,” so the 1d interval has surely to do with the current FSRS parameters. (FSRS has learned that my forgetting a new card is a rather dire event and that I shall review it again the day after.)

If the algorithm has learned my behaviour without serious downsides, I think the best I can do is keep everything as is, which I rather enjoy anyway.

I guess my final question is: are there potential downsides to maintaining my current approach that I should be aware of? Can FSRS properly distinguish between new and older cards in its algorithm?

4

Are you sure you changed your steps for all of your decks/subdecks? If a deck is using a different Deck Options preset, it might be using different steps.

Without looking at your parameters, it’s hard to say whether or not they could be the reason. [If you want to ask about your specific parameters, please post them, as text.]

I’m going to wait to answer these questions until we figure out what is going on.

5

I don’t think anything wrong is going on… I actually find it kind of obvious now.

  1. No, I didn’t change the params of all decks… just of that one deck.
  2. I did change it, as proved by the change in behaviour.
  3. Having a single step by 10 minutes means when learning a new card for the first time I’m presented with two options: clicking “good” and seeing it in ~5 days (without any short step at all) or clicking “again” and see it in 10 minutes, but lowering the ease of the card, leading to the next “good” interval proposed by the algorithm to be 1 day. I think it makes sense for good → good to have a different proposed interval time than again → good?
    My learning and relearning params are simply (now back to) “1m 10m”. I experimented using only “10m” as a learning step while retaining “1m 10m” as relearning steps (If I lapse on a review card I’d rather see it immediately again).
6

You’re using FSRS, so that’s definitely not what’s happening. The reason the post-Learn interval is changing is based on the initial Stability set based on the first grade you give it. Still haven’t seen your parameters, so I can’t tell you more than that.

7

I think it’s only natural for the algorithm to make a distinction between good → good and again → good, I did use them differently after all… I don’t see how anything else but FSRS parameters could explain this.
I used the term “ease” albeit maybe official FSRS terminology uses the term difficulty or some other term? I think you might be nitpicking a bit.

Here are my FSRS parameters for the deck:
0.0100, 0.0354, 3.3659, 70.1346, 6.9922, 0.6424, 2.0011, 0.0010, 1.6801, 0.0015, 1.1550, 1.8510, 0.1759, 0.2169, 2.2047, 0.5887, 3.5573, 0.6986, 0.4657

I’d be very thankful to you if you could answer these:

I guess my final question is: are there potential downsides to maintaining my current approach that I should be aware of? Can FSRS properly distinguish between new and older cards in its algorithm?

8

The default algorithm has a factor called “Ease” – which is handled completely differently. But Difficulty isn’t the reason either. I explained how the parameters factor into that.

Well, that’s concerning. The first 4 parameters are your initial Stabilities for each grade. So if your initial grade is Again or Hard, you’re starting basically at 0d Stability, for an initial-Good 3d Stability, and for an initial-Easy 70d Stability. If you told me you hardly ever use Hard and Easy, that’s probably okay, but otherwise, that’s pretty concerning. In the past, have you mis-used Hard for incorrect answers (when you should have used Again)?

Your system essentially inflates the grades of your New cards. You’re passing cards based on the length of the delay you want, instead of based on the honest and accurate grade. That’s the kind of thing we’re used to seeing with Hard-misuse (where it hits initial-S for Good and Easy) – but yours might be more of a “Good-misuse” situation (hitting only initial-S for Easy). I think that’s what is causing that huge jump in your Stability.

Yes. I’m not sure exactly what you’re trying to assess with that question, but yes, FSRS does that.

9

Thank you.

I don’t ever click Hard.
I only did it 11 times on a total of 10000 reviews, 7 times out of 11 on new cards when I was starting out.
I’ve very sparsely clicked Easy on young cards (0.67%) and mature cards (2.85%) and I’ve since decided to stop using it for those.
I do occasionally use easy for new cards though, to signal I already knew them very well beforehand (Easy answers make up 8.24% for new cards, 5.54% in the last month).
I might have been influenced by the interval times in the past but I’ve since disabled them, and I don’t think they’ve ever influenced me to click Easy on new cards, in fact I expect the initial-S for Easy to continue to grow as months go by (because when I use Easy I’m confident I’ll remember a card in 3-4 months, even more for some, not just 70 days).
I’ve never misused Hard for Again, so I think all this matches the way I use the buttons (this is also confirmed by the 3.86% RMSE I think).

If you don’t have any further observations, I’d call this solved.
Thank you again.

10

It sounds like those parameters are working for you then! That’s great.

1 Like
11

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.