As the title suggests, I’m trying to find the most recommended Display Order for FSRS. I found different ones from the Anking, Reddit posts, and also Xelieu’s Lazy Guide, and dont know which is the “appropriate” one for FSRS.
If you’re looking for review sort order, easy cards first is best in terms of efficiencydescending retrievability in terms of maintaining the desired retention.
However, this only applies when you have a backlog. Otherwise, just set it to random.
I just remembered their was the new chart with FSRS-6:
So at least this is showing that ascending retrivability gets the best efficiency. So the opposite of the other sort order I mentioned.
4 Likes
If you are studying your cards when they are due, there is no “most recommended” Display Order. FSRS doesn’t need you to study your cards in any particular order.
For overdue / backlog cards – irrespective of FSRS – you may want to choose an order that maximizes your chance of success, however you define that. The table that sorata posted always looks like gobbledygook to me, and it’s somewhat useless without any explanation – but the overall takeaway is: what order you choose depends on what you want to prioritize. For instance, Descending Retrievability gives the highest average retention (94.9%), which is why that is routinely recommended for backlog studying.
4 Likes
Thank you so much for your answers!
I’ll go with “Descending retrievability” as the “Review sort order” value, since it gives the highest retention, and also its the most recommended one.
But I still don’t know what to use for “New/review order” and “Interday learning/review order“ (I don’t want to change the “New card gather order” and “New card sort order“ options).
To provide some examples, Anking recommends “New/review order: Show after reviews“ and “Interday learning/review order: Show before reviews“, and the Xelieu’s Lazy Guide recommends “New/review order: Show before reviews“ and “Interday learning/review order: Mix with reviews“.
There’s no one-size-fits-all answer for that either. [Well, there sort of is for “interday.” It doesn’t matter unless you have steps that can’t be finished in a day (long steps, or a lot of steps). Both of those things are recommended against, so you can leave that option alone.]
So – when do you want to study the New cards that are being introduced today? Pick what you like. And if it doesn’t make you happy, pick something else tomorrow.
1 Like
I’ve been thinking about this. A generally fine default approach would be to go by ascending retrievability. A slightly better approach might be to go by ascending order of retrievability, but not evaluated at the time of the review but 24hr in the future of the review (i.e. so that you minimize the amount of R loss between the time of the review and the next day). However, cards with very low retrievability (i.e. t >> S, R<<1%), it seems that there is not even any point in doing them and that they should just be re-marked as new, and new cards, of course, should be limited until the backlog is cleared out. Conversely, for cards with low R (1%-10%), the act of getting them correct would increase their S by so much that it seems almost a waste not do them in the very near future.
In the end, I think the gain/loss in R in any strategy is statistically negligible as long as the amount of time that the backlog is allowed to exist is much shorter than the S for the cards.
If the user is sufficiently motivated, then he can even do new/young cards first before clearing out the old cards, as this will not stop his progress of adding in new cards and promoting them to older ones, and the difference in R at the time of review vs. DR will be negligible.
Whatever strategy gets the user to clear it out as quickly as possible is the best. Probably the ordering is not the problem with this as much as the mental state of the user is.
Wouldn’t it depend on Optimal retention and Desired retention? If Optimal retention = 80% and your Desired retention = 95%, then Ascending retention is likely to shift average_true_retention towards Optimal retention. If your Optimal retention is 80% and Desired retention is 80%, then Descending retention will allow average_true_retention to be closer to Optimal retention.
Therefore, with a small backlog, I would not do anything. If I had to return to anki a year later, I would set Desired retention = Optimal retention and enable Reschedule cards on change (preferably using FSRS helper). Review sort order = Descending retrievability
1 Like
We called this PRL and tested this, doesn’t work well. PRL = Potential Retreivability Loss.
Also, we tried PSG, which is Potential Stability Gain. This didn’t work as expected either.
I strongly suspect the reason the testings were unintuitive is because we missed something, i.e. something was wrong with the simulations, but I can’t say for sure. Take a look at this thread if you’re interested: New Sort Oder: PSG
The linked discussion thread there has too many replies so you might want to sort by likes, and read the top ones.
1 Like
That thread is over a year old, so forgive me for commenting here.
If I’m not mistaken in my mathematics, then
PSG = (S_recall ÷ S) × R
should more accurately be
PSG = (S_recall - S) × R
If I’m not mistaken. Technically it should be the integral of the forgetting curve from that point in time to infinity, but SxR is probably roughly proportional to that anyway, so it’s probably close enough. But I dunno, it might not apply when dealing with cards like this. Integrating the forgetting curve is the correct method.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.