Definitely the best solution. I really don’t like just getting rid of some of the more inefficient sorts because there are good reasons to use them sometimes.
For example, Order Added is probably one of the worst sorts in the simulations. I use that for a deck of Chinese character flashcards where I practice writing them. I created the flashcards in order of how frequently each character is used in the real world. When I get a backlog, I kinda don’t care about getting through them efficiently, I want to prioritize studying the most frequently used characters first.
You can see in this graph how noisy difficulty_asc is though. I’m wondering if the average retrievability for that is also noisy and the simulation just happened to end when it was in a good place. Maybe if you did 25k cards instead of 20k, it wouldn’t look as good. It might just be kinda random because it’s fluctuating. Just a hypothesis, I’m not sure if this graph connects to retrievability in that way.
Because when you get a card wrong today, the next time it shows up, it’s going to be at the bottom of the pile. You’re constantly focusing on the cards you’ve forgotten the most, which means those cards that are freshly due are waiting and rapidly losing Retrievability. You’re going to forget them too by the time you finally get to them. It’s causing you to study cards way more frequently than you should be.
@expertium I still haven’t read everything here or that one link you provided, so confirm if my understanding is wrong. Can you do a simulation where the user with a backlog studies all prop:due=0 cards everyday and his backlog is being covered seperately?
This is the method manual recommends so it’ll be useful to see what sort order to use here. FYI we still have a detailed section of “how to handle backlog” as pending if we’re meant to do it eventually. I’m trying to understand what a user with backlog should do.
I think it was already taken into consideration that this was to be the case if the user has a backlog AND the user can only do a limited number of reviews per day (e.g. in my case I have a backlog of 20000 cards and can only do 1500 per day).
If that is not the case, then the sorting order doesn’t really have a notable effect. I have spent the entire time reviewing my backlog carts under the current Relative Overdueness sorting (Retrievability Ascending), and it was suboptimal to say the least with such a massive backlog like mine.
I will now try Difficulty Ascending for a change and await the coming of Reverse Relative Overdueness (or Retrievability Descending) - I believe this addition should be simple
In the simulation, there is a fixed number of new cards (20 per day) and reviews (80 per day). If the user has more than 80 due cards, he will only review 80.
This sparked a question in my head. What usually happens for me is that the cards that I get wrong dont get send to the back of the pile the next day but instead remain at the front so I am in constantly in a loop. Either way the end result is the same.
But this: Does Difficulty change with the number of days passed without review like Retrievability (R value) does @Expertium
If so, how robust is Difficulty as a sorting criteria to tackle huge backlogs where it could take lots and lots of days
Yes, pretty sure they’re specifically setting a limit for the cards/day being studied so that there will be a backlog and we can see the simulation take that into account.
Just to elaborate on this side note, because now I’m realizing it’s totally feasible, you wouldn’t actually have to calculate a derivative. It would basically be:
Potential Retrievability Loss (PRL) = R(Today) - R(Tomorrow)
“PRL Descending” would be the best sort. You want to be studying those cards that are going to lose the most retrievability if you don’t study them today. It’s close to what Retrievability Descending is doing, but it’s a little better on certain edge cases.
Pretty sure I can figure out how to add that to the simulations, so I’ll do that when I have time if you guys don’t want to get to it first.
You might get a slightly different sort here. edit: okay no ignore this. I can’t wrap my head around it.
I like your idea actually. Overdueness sorts cards on the basis of “how much recall-ability it has lost” and here it is “how much recall-ability it’s going to lose in another day.” Though you might wanna chnage the formulae.
@Expertium Is this possible? I wonder if this can become the holy grail of managing backlogs. Also, whether this will work for really high stabilities when R won’t show much change in a short span of time.
Edit: Actually, given the shape of forgetting curve shouldn’t the proposal here be pretty much R descending? Not sure now how useful the PRL calculation will be. Try it out still, Richard!
