No. Total number of remembered cards isn’t what you think it is.
total_remembered = int(card["retrievability"].sum())
In plain English, it’s the sum of probabilities of recall across all cards that you have reviewed at least once. Higher = better.
Sooooooo…is Difficulty Ascending the best option for huge backlogs… 
Either that or Retrievability Descending, but we don’t have the latter in Anki at the moment
1 Like
Then total time spent is definitely the better metric. It’s hard to even parse the meaning of Time per Remembered Card. Retrievability Descending has the lowest total time spent.
The sum of retrievabilities is actually quite simple to interpret. It’s just the number of cards that you are expected to recall if you review all those cards right now.
So if you have 100 cards and the sum of retrievabilities is 95, then you are expected to recall 95 of those cards if you were to review them right now.
But that changes after every review, so are you just taking the final state, or you’re factoring in Retrievability at the time of each review somehow? Seems like you’re just looking at the final state, which doesn’t seem useful at first glance.
The final state at the end of the simulation.
Here’s what I’ve been doing:
- Make a custom deck and include
prop:r<0.9 prop:r>0.85in the search. - Ordering really doesn’t matter, but I use Relative Overdueness.
- When you go through those, change
prop:r>0.85toprop:r>0.80, thenprop:r>0.75(if that adds too many, I’ll make the increment smaller and go toprop:r>0.74or something), etc. - I just keep working my way down. The next day, when new cards come due, they will automatically be included because their retrievability is above that threshold.
I’m basically sorting by Retrievability Descending with a lot of extra steps.
Daaaaaaaaamn I just realized, looking at the simulation, that it is much, much better than the current sorting by Relative Overdueness. For some reason, I thought it was intuitive you want to review cards you are about to forget - that is the point of Anki. But this is a real eye-opener.
1 Like
That’s not what Relative Overdueness does anyway. It’s making you review the cards you’ve forgotten the most. If anything, Reverse Relative Overdueness is having you review the ones you’re about to forget.
You can create a filtered deck with Relative Overdueness order. By reducing the filtered deck limit, you will send the cards with the highest Retrievability to the parent deck.
1 Like
So basically you were looking to maximize average retrievability after a set number of learned cards. Yeah, I feel like minimizing time spent is more important.
How about we create different sort orders and test them and keep the best?
I see what you’re saying. That’s not a bad idea. Accomplishes the same thing.
@Expertium Can you ask them for these charts, but with the other sorts added since they did more? Or at least with the couple best sorts just to compare them.
Interesting to see, sorting by Relative Overdueness isn’t any better than sorting by straight up Random.
@L.M.Sherlock please make these charts for stability ascending + stability descending
Feel like throwing some randomized lengths of skipped days in there would be good too.
Looks like they already did a month ago.
Where?