Usually, all other things being equal, random is used.
Random is also one of the worst sorts when you have a backlog.
I could see primary sort being difficulty_asc with secondary sort being retrievability_desc being really good. Maybe this is what PSG is getting close to doing anyway.
Another solution would be to stop Difficulty from reaching a perfect 10 so easily. Implement some kind of asymptote so cards never actually reach 10 but approach it very slowly.
Now I’m curious what the sim is defaulting to when there are ties, what’s its secondary sort?
Already. The more your difficulty differs from the basic one (good), the stronger the return to the average is.
There has been a change in FSRS 5 that I like, although we don’t know if it’s for the better or for the worse. The difficulty now tends to the basic (easy).
The problem is that the optimizer often decides which is better for you w7=0.
It’s extremely slow if you are answering Good though. For me, I can answer Good 100 times in a row, then one answer of Again puts me right back at D=10. So basically most the cards that I answered Again at their last review are tied at D=10.
If there was an asymptote, then D would more truly represent the cards you’ve answered Again or Hard the most on in their history, which is what it’s basically telling you now but with a hard cap.
The problem is that D is needed to calculate the optimal interval. And as it is now, it does its job better.
The fact that it can be used to better deal with the backlog is just luck.
Might be even betterier with an asymptote
Really don’t think so. I thought this too, and I changed all sorts of things in the sim to try to find where it wouldn’t keep doing so well and it always outperformed expectations. There’s some reason it works, I just don’t fully get it.
We don’t know if it works well in other cases.
Yeah, we do random() for ties. This was talked here before but I guess with over 400 replies it’s normal to forget.
Then you’re saying something else. You do not distrust the metric per se, you just want the sim to run even after all cards are cleared. Sure, we can run sim longer and result is basically the same.
Sorry, I wasn’t careful before and messed up with the formula.
Does the same thing.
PSG = (S_recall ÷ S) × R
If you substitute the base weights into the formula and see how the value changes, you will get
PSG_desc = (S_recall ÷ S) × R = difficulty_asc x stability_asc x retrievability_asc
However, at very low R values, retrievability_asc turns into retrievability_desc
PSG_desc is just a worse version of difficulty_asc
We already know that stability_asc and retrievability_asc are bad sorts.
I am not sure about this. How did you get stability_asc in the formula for PSG itself? Is there a 1/S in the formula? edit: okay there is but it cancels out.
And you looked at the metrics for that. And metrics show PSG performs well. Even better for total_learned.
On a different note, we’re still fine with PSG right? It’s an improvement over what was already suggested.
@L.M.Sherlock please add PSG here, I want to test it
I see that you have already figured it out, but still:
And as far as we know, it’s not very good.
It works well only because it’s pretty much a difficulty_asc but worse.
Maybe I’m wrong and there is something in this, but in this case it is necessary to find it and discard all unnecessary things.
But you don’t understand why. The only explanation I see is because there is (11-D) and it gives us a difficulty_asc.
Have you already checked something like this?
Descending
(11-D)*R
(11-D)*S
(11-D)*S*R
Or an attempt to choose so that some parameters do not affect too much or too little
(11-D)*S^(0.5)*R^2
No, even then it’s useless. I can’t think of a case where the metric is useful.
It’s basically just telling you the mean of the distribution (the sum, but divide the sum and you get the mean) of retrievability. Even if you run the sim until all cards are in prop:r>0.9, it’s still just telling you the mean of the distribution of R above 0.9.
Why do we care about that? How is that actionable?
What’s the objective when you’re studying with Anki?
- To maximize the mean of the R distribution after a set number of hours/cards studied, regardless of the shape of the distribution?
- To get all cards above your set desired retention?
- Something else?
To me, it’s number 2, and then you’d judge sorting methods based on how long or how many cards studied it takes to get there. If you want to maximize average retention, set your DR higher.
The mean isn’t just a mathematical mumbo jumbo. It’s giving you the number with which you can replace all the other numbers. That’s what we are trying to do every time we calculated averages.
It’s useful because now you know how much of the collection you know and how much you have forgotten.