With FSRS 5, Allowing new Reviews before next day to increase average retention

Anki FSRS
1

The “Desired Retention” is a threshold evaluated card by card. If your deck is composed of 10 cards, and 5 of them have their “Estimated Retention” lower than your Threshold (let say 80%), then it will be gathered today.

But, it means that by definition, each card has an estimated retention LOWER than 80%. Worse, if you have a lot of low intervals/stability compared to your high intervals/stability, some of those cards might already have a retention way lower than 80%, it might already have dropped to 20-50%, for example.

So, it is perfectly normal that “Average Review Retention” is lower than “Expected Individual Retention”. The only way to compensate that would for Anki to give you already some reviews of some cards above your Desired Retention to compensate the very low one. But it start to defeat its purpose, and it could lead to a lot of 90% retention reviews because one already has 10% retention for that day.

It does however mean that if you want to have at least 80% retention, you should put the desired retention HIGHER than 80%. How much higher ? The more “short intervals” you have compared to the “longer intervals”, the higher !

To make it easy to understand, imagine putting FSRS desired retentino to 99%, but having a lot of cards you just learnt today. Even if FSRS would be extremely pessimistic, it will gather those cards only once per day. To get that 99%, you would probably have to have shorter intervals than 24h.

A potential improvement of Anki could be to allow such <24 reviews, now that FSRS5 (Anki 24.10RC) already compute same-day reviews for FSRS optimization. In this way, you would not have to wait for the estimated retention to drop to 20-30% before having to redo that card. You could just do it 4-5 times one day.

Of course, it’s not proven that it would help increase how fast stability will increase. But it would definitely help getting a desired retention closer than the one you get when doing reviews

2

Have you tried “Advance” in FSRS Helper?

3

You are misusing some terminology, making up other terminology, and I think misunderstanding how FSRS does what it does. It’s hard to know where to begin to correct some of your statements.

2 Likes
4

Maybe I’m lost, but this sounds like a bit of a contradiction.

Since FSRS5 allows for control of learning steps, including same-day intervals, does that not already allow for multiple <24-hour revisions?

5

Well unfortunately this is not what I’ve observed, let me take this card for example :

image
image1712×2180 296 KB

It took me 4 tries over 23 minutes to get it right. The Stability was computed to 37 minutes (a.k.a 37 min to go from 100% to 90%), so if my desired Retention was 90%, I would already benefit from reviewing it after those 37 min.

image
image750×884 38.4 KB
]

Instead, it is now 6AM and my Retrievability is now only 39%

image
image1622×960 52.5 KB

So while the latest RC takes indeed into account same-day review to compute the stability, you still “succeed it once”.

In my case, having dropped the desired retention from 80% to 75%, my True Retention went from ~75% to ~62%

image
image1298×372 17.7 KB

I wondered why, but after some analysis I found out that by dropping the desired retention, the ratio “new cards/day”/“total review” being higher, I have a lot of lot of low stability cards in my reviews.

Over my 191 reviews of today, I have 93 cards with stability lower than 24h (and thus 98 greater).

Note this is not only applying to completely new cards, this one for example has a R of 40-50% still after 7 days and could probably benefit from more regular review

image
image1616×1980 441 KB

Note, @Aleksej, that “Advance” undue cards would mean surely getting those earlier, but it would also mean getting cards from tomorrow that still have Retrievability > Desired Retention, so it’s unfortunately not really fixing the issue. “Review Forgotten Cards” would also not really help because you could have a card just out from Learning stage that is not a Forgotten one, but still with maybe a very low stability and thus benefitting from a second same-day review

@Danika_Dakika : To make it simpler for you : If a card is due if only Retrievability < Desired Retention, then it’s not surprise that actual retention is lower than the desired. That check (R < DR) is on a card-by-card basis, while the “True Retention” (TR) is a score averaged on the result of that day. But if all those cards had Retrievability way lower (because short intervals, lot of news/day…), then the TR will be way lower. If you see any contradiction, just take one, refuting an hypothesis only requires one.

6

@sound, to get the desired behaviour, execute mw.col._set_enable_fsrs_short_term_with_steps(True) in the Debug Console.

Note: This option is in the Debug Console because it is still in the experimental stage.

2 Likes
7

Retrievability is a prediction of a probability. Retention is an outcome. You are falsely equating them.

1 Like
8

Thanks for the pointer ! Found out this topic when searching on that option : FSRS 5: <1d Scheduling and Learning Steps - #124 by DerIshmaelite

I’ll test it to see how good or bad things goes with my current setup :slight_smile:

9

I tested the feature, but I see now some cards seems to be blocked in Relearning even though I succeeded them twice

image
image892×1224 90.2 KB

Checked a bit the code here anki/rslib/src/scheduler/states/relearning.rs at e453e02247feefb902b8cfb58fbff8c915b60d52 · L-M-Sherlock/anki · GitHub

If I understand correctly @L.M.Sherlock , to go back from “relearn” to “review”, the interval need to be >0.5(d?)

So I guess even if I had a Interval 1d in the Card info, it might not really be what was compared to that threshold, but the state.good.interval ?

I’ll check tomorrow how they behave once their interval is bigger than half a day :slight_smile: (12h then I guess)

10

Just to be sure, do you have FSRS 5 parameters (19 values) in your Deck Options?

If you clicked Optimize, but you still have the older 17 parameters, then you might want to clear the older parameters and click Optimize again. Wanting to try out the FSRS short term scheduler is probably the only case where I would suggest deleting the FSRS 4.5 parameters for generating the FSRS 5 parameters.

11

Hello ! Yes, I do have 19 values indeed :
“0.0647, 0.2208, 1.3910, 16.2874, 7.3853, 0.9048, 2.0104, 0.0010, 1.1176, 0.2663, 0.6512, 1.8657, 0.1284, 0.1972, 2.4391, 0.0009, 3.5759, 0.8641, 0.9777”

I continued testing and everything seems fine once the interval is big enough again :slight_smile:

image
image1720×1810 400 KB