Allowing more decimals for Desired Retention closer to 100%

[sound]

Hello,

I have an alphabet deck I use for my kanas, and since intervals really increase drastically closer to 100%, I was wondering how feasible it would be to allow more decimals than 2 when choosing a Desired Retention ?

For examplea 100 day interval for .90, for .98 is 18day and for .99 is 9day.

Allowing a .985, .995, could be very good for decks like that.

But of course, I don’t know how much decimals FSRS has for Retrievability, so it might no be doable.

[Danika_Dakika]

98%-99% is extremely high for a Desired Retention – that should hardly ever be necessary. Why are you wanting to study the same material so frequently?

The “A 100 day interval will become …” message doesn’t mean that your cards will be given 100-day intervals – , it just gives you a scale for comparison on how changing DR will change your intervals.

[Expertium]

Personally, I’m against it. I don’t see value in this kind of micro-tuning.
Plus, 99% is already extremely high. Originally, desired retention only went as high as 97%, but I guess some people just love cramming, so they requested to change the cap. At that point it’s basically massed repetition and not spaced repetition.

[sound]

It is still spaced, but for decks like learning foreign alphabets it’s what allow you to have every day ~30-40 of the most problematic letters you have while the one you never fail will still have a 10+ day interval.

Problem is that .99 is basically twice the interval than .98 and the very last option.

.98 curve

.99 curve

a .995, .998, would be nice step forward.

For knowledge domains with ambiguous information but low total volume of cards, with high level of precision needed, this is perfectly reasonable. What is the chance of you missing a letter in your mother tongue? Pretty small, right. What level of error can we expect someone to miss an element of the periodic table ? Almost none.

A pure mass repetition system would make me cycle through all 227 cards every day, when in fact only 30-40 still need a 1-2d interval. Expecting someone to go through the full alphabet/full periodic table each time they want to review instead of increasing by smaller and smaller step their accuracy is like telling them to cycle through the full alphabet or periodic table if there are only 3-4 that they keep getting wrong.

[Expertium]

Oh man, I can list quite a few reasons why it’s not a good idea:

1. I’m not sure that FSRS can predict R with that level of accuracy. In other words, I don’t think FSRS can distinguish a card that has a 99.5% chance of being recalled from a card that has a 99.6% chance of being recalled.
2. The main indicator of whether FSRS is doing its job is True Retention. However, at such high levels of desired retention, you’d need an astronomical number of reviews to distinguish between, say, 99.8% retention and 99.9% retention. I ran a simple simulation in Excel: even with ten thousand reviews it’s not guaranteed that there will be a statistically significant difference between a schedule with 99.8% target retention and a schedule with 99.9% target retention. Simply put, if I give you 10,000 data points from the 99.9% schedule and 10,000 data points from the 99.8% schedule and ask you which is which, you won’t be able to tell.
3. At 99.5%, let alone 99.9% retention, the intervals will be pretty much static with almost no growth. With default FSRS parameters and 99.5% desired retention, do you know how many consecutive "Good"s you’d need to make your interval 10 days? Thirty-seven. That’s right, 37 "Good"s in a row (no Again, Hard or Easy) just to have a 10-day interval. I’m serious when I say this is no longer spaced repetition.
4. It will make users more concerned about micro-tweaking their retention and they will be like “Hmmm, my desired retention is 89.5%, I think I should raise it to 89.6%”.

4 is pretty weak, but 1-3 are very strong objections.
FSRS probably cannot distinguish between 99.8% and 99.9% p(recall) cards, or between 99.5% and 99.6% p(recall) cards.
Even if it can, you won’t be able to tell the difference anyway.
Even if you can tell the difference, you have effectively un-Ankied Anki. I mean, 37 consecutive "Good"s to get to 10-day intervals, for Christ’s sake!

[Danika_Dakika]

@sound

it’s what allow you to have every day ~30-40 of the most problematic letters you have

Easier than torturing your whole deck:
Filtered Decks - Anki Manual + Searching - Anki Manual

How about something like deck:deckname prop:lapses>3 prop:s<5 ?