In what ways does FSRS struggle with lots of learning steps or longer learning steps?

Anki FSRS
1

I often see it recommended that you should only use a small number of learning steps with FSRS, and that the steps should be under 24 hours. The most common reason given in online discussions seems to be “it messes with the algorithm”, without any explanation of exactly how it does that.

The Anki manual gives a little more detail, but most of those reasons don’t sound like “it messes with the algorithm”:

  1. Learning steps aren’t needed to protect cards from “ease hell” with FSRS the way some people thought they needed to with SM-2.
  2. If you keep learning steps short, FSRS can schedule your cards at a time it calculates is optimal for you.
  3. When you transition from learning to FSRS-controlled scheduling, the “Hard” button (staying in learning) may show a longer amount of time than the “Good” button (FSRS scheduled).
  4. Repeating a card multiple times a day may not contribute to long-term memory.

Of the reasons given by that section of the manual, only #3 sounds like “it messes with the algorithm” (and only in a minorly inconvenient one-off way). The others are more “reasons why it might be optimal not to”, rather than “this specific bad thing might happen if you do”.

Elsewhere it mentions that FSRS doesn’t use Anki’s distinction between “review card” and “learning card”, and will decide which formula to apply based on interval < 1 day or not. (Which may be exacerbated by 8-hour intervals becoming 1-day intervals if they happen to cross an Anki day boundary.) That sounds like it could have higher “messes with the algorithm” potential (but only if FSRS’s formula for intervals longer than 1 day is inappropriate for learning that needs intervals longer than 1 day).

I suspect there is more “it messes with the algorithm” going on. In particular, I suspect there is something going on with stability, since I have seen cards that have had some 12-24 hour intervals during learning graduate with intervals like 9-14 days. (I haven’t fully traced how this happens, but it seems to relate to the initial stability the card starts with, and how the stability grows from there.)

Before I spend a lot of time trying to figure this out for myself, what do people more familiar with the algorithm already know?

If a user has either lots of learning steps and/or longer learning steps before the card is finally handed off to FSRS for scheduling, what are the specific ways that might mess with the algorithm?

(To be clear, this is not a question about how people should do learning steps with FSRS. That has already been asked and answered many times. My question “why?”. What are the specific ways in which having lots of learning steps and/or longer learning steps messes with FSRS?)

2

I don’t think it’s necessarily accurate to say that FSRS “struggles” in those situations or that they “mess with” the algorithm.

The reason to use shorter steps is the clearest. Using longer steps interferes with FSRS taking over and scheduling the card. If everything is working correctly, FSRS will do a better job scheduling a card for 1d+ intervals than you would do by trying to control that with fixed steps. [If everything is not working correctly, it’s better to fix how FSRS is working for you, than to push FSRS out of the way and do it yourself.]

As far as the reason to use fewer steps – this recently came up in another thread, so I’ll quote myself as a jumping-off point –

Why one step though?

I make that recommendation based (1) on FSRS research that found multiple same-day steps have little effect on long-term retention, and (2) on FSRS functionality that considers multiple grades in Learn to set the first interval when the card graduates to Review.

For some users, too many correct grades in Learn can cause a drastic increase in that post-Learn interval. All of the analysis of FSRS points to that effect being correct, based on those users’ review histories (meaning, it’s not a “bug”[1]) – but that doesn’t make it look any more rational when they see it.

A single step, as a starting point, deals with both points. Some users may find they need to add an additional step, shorter or longer, to deal with issues they are having in Learn, but that can’t be known until they get started.

  1. Is it part of FSRS not having a model for near-term memory? Does that make it so FSRS isn’t able to consider those near-term/same-day grades appropriately? I’ve asked those questions and been told that isn’t an issue. ↩︎