Expertium, is D value capped at 100% in the sim?
Also, I can’t see anything here. It seems difficulty_asc isn’t that good when you stopped learning new cards (drops in retention mid-way). I was planning to use this as a sort order but now I don’t know.
Could you not increase the DR? The retention is still .8 I’m assuming?
Actually, now that we’re planning to change the default sort orders it might make sense to run the simulation with default settings. That is, 20 new cards/day, 200 review limit, DR = .9 and default params. Can anyone do that? Richard?
Of course it’s capped.
If its possible to do it without the cap, can you do it? If that works still better, Anki can internally use it for cards with 100% D. The users won’t notice it. And of course, it’s useful because some users have most of their cards in the high difficulty region.
But that’s not how FSRS works. Also, I can’t think of a way to have difficulty capped inside FSRS and uncapped for the purposes of sorting.
You mean in the simulation? That I don’t know, I haven’t even visited that GH repo you linked to. I’m betting you can do it.
For Anki, we can possibly store D value uncapped but show people 100% when D value is above it. Visual hack.
I thought the reason this was made to cap at 1 is dae didn’t want it to get confusing for users? If that’s true and uncapping D value doesn’t have any side-effects, why was it capped in the first place? I remember you told me D was meant to be uncapped in FSRS. Or I’m misremembering?
No, D is capped in FSRS, it’s just that in FSRS it’s between 1 and 10, not between 0 and 1. The conversion to percentages is unnecessary, and it was a bad decision. It can make people think that D is a probability (it’s not).
Oh, okay I misunderstood you.
@Expertium I only have one request. Can you make a table like Sherlock did? Because that graph is very much incomprehensible right now. Actually, two tables. One for when you are learning new cards everyday. One where you are just reviewing old material.
Yeah, I’ll run it. Gonna let it run while I’m at work, so you won’t hear back for a while.
Make tables lol. Those graphs are un-understandable.
Oh, wait what? PRL is best in optimising time spent per card? Zamn.
add_order_desc being so high up is pretty surprising. I’m also continuously surprised by how well difficulty_asc does. Maybe I just don’t understand how difficulty is calculated well enough, or maybe that metric doesn’t make sense for my deck because of how big my backlog is.
This is my entire collection. With the vast majority being over 95%, I just don’t see how sorting by that would be that useful. But maybe that would change if I was studying more consistently.
It’s certainly not applicable for everyone. Are you doing the simulation using your collection? I assume not.
If you don’t mind, two tables would be better:
I sorted that by total time, add_order_desc was technically the best in that time per card category, but that category is basically (average retrievability)/time. The way I’m running this sim is to stop when they are all above the DR, so the average retrievability isn’t really as useful, so the time per isn’t either.
No, I’m running this straight off of @L.M.Sherlock’s code and just modified it.
I think you should just take a large enough number like Expertium and be done. The day when prop:due=0 returns 0 cards is a bit arbitrary.
The other change I want to make is to see total cards reviewed. I’m skeptical about the way time is being applied here (I think it’s by difficulty), or at least I don’t think that applies to the way I study at all. Might work well for the average user.
But for example, I have Chinese vocab cards that have an example sentence on each. I spend time to read the example sentence out loud so I’m practicing speaking as well. Those cards take much longer even though they’re fairly easy for me. I think total cards studied would give me more information than a simulated time spent.
