Can you do this too please? Ordering Request: Reverse Relative Overdueness - #112 by sorata
I’m thinking if a user has a backlog one thing we ask them to do is create a seperate filtered deck for overdue cards and keep doing all overdue cards after doing prop:due=0 cards. I’m inquiring what sort order should be used in that filtered deck.
It requires a completely different framework of simulator.
Besides, if the user studies all prop:due=0 cards everyday, where does the backlog come from? I mean, how to make a backlog in the simulator?
Start the simulation with an existing backlog.
It’s not an easy task, because we need to assign the stability, difficulty, interval, due date and a dozen of properties for each card of the backlog. These properties may have different impact on different sorting methods.
I imagined a backlog could be created in simulation, and then you’d take that and change the rules mid-way/start anew.
Anyways, do try out OP’s suggestion.
At this point the graphs have so many curves that it’s impossible to tell what’s going on. Please add a table with the following numbers:
Yes, I am well aware of that strategy. It would be a huge pain to make a new deck for the stale cards and then return them to their main deck. They will lose the original deck they are assigned to if the main deck has many subdecks…So I figured making a new sort order is a good reversible solution.
But idk, maybe a far more robust sorting method like reverse retrievability is a better option. After all it is the simulations that count.
Also, do the results of these sorting method simulations remain consistent regardless of how big the backlog is 
For example, would Difficulty Ascending become worse off with increasing backlog size provided that Difficulty doesn’t change with the number of days passed, which are required to work through the backlog like Retrievability does
Because we need to find out which sort orders allow the user to maintain retention at the desired level even when he has a backlog. Also, rename “seconds_per_memorization” to “seconds_per_remembered card” and use only 1 digit after the decimal place. And remove the index numbers (leftmost column)
For average true retention I suggest using 3 digits after the decimal place
It’s interesting that Retrievability Descending is more time-efficient than PRL
I ran this on my PC
Another interesting discovery - sorting by stability works well enough either way, but ascending is better for maintaining retention
Also, Jarrett, another suggestion: learn_span = deck_size / learn_limit_perday. This way the number of days to simulate is automatically adjusted if you change the deck size from 20k ot something else
I’m surprised PRL didn’t surpass R descending.
@rich70521 do you have ideas for a shorter, easier to understand name than “Reverse Relative Overdueness”? I suggested “Least Overdue First”, but Dae said it doesn’t convey that it’s relative
Btw, I wonder if these results are robust to fuzz. Try adding fuzz
As a note regarding the comments above about the user interface and new users I came across this feedback on another application:
“thanks for this feedback. We haven’t decided whether to build this yet. We’re focusing on improving our own SRS flashcards first, for example from next week you’ll be able to import vocab from elsewhere, and edit words and translations you’ve saved. I understand Anki provides lots of flexibility, though I personally tried it and found the design to be unintuitive, so we’re trying to provide something better.”
I think that the idea of a user interface that has 2 modes ‘beginner’ and ‘advanced’ would help prevent such a perception (along with other measures), and the beginner mode should use more task orientated terminology as was suggested earlier in the thread, eg. the user could choose scheduling from a list like “normal study”, “backlog” rather than being expected to understand particular algorithms.
Doing some quick math, the derivative of R comes out to be proportional to R3/S. For the reason mentioned by you, I think that sorting in the descending order of R3/S can be quite effective (when there is a backlog).
It makes no sense to analyze messages that are used to promote another application at the expense of anki.
And backlog size as well…
