Allow configuring DR for each deck in a preset

See how in “Daily limits” it’s possible to configure limits for the preset and for individual decks?

I think it would be nice to do that with desired retention too, like this:

This way, if someone wants to have the same parameters for several decks but different DR for some decks, they can configure it without making multiple presets.

6 Likes

I strongly support this idea!

Wouldn’t optimizing become harder then?

Why would it?

When using the “save to all subdecks” feature to change new ones per day, users may get unexpected results.

1 Like

How does the “This Deck” Option work by the way? Does it internally create a child-preset which inherits from its parent-preset?

Optimizing of parameters is dependent on the DR, no? If the DR for each deck is different, then how will it choose which DR to use to optimize?

Nope. The opposite.

2 Likes

FSRS params are optimised on the reviews and the grades. DR is just deciding when your cards get due. FYI the default params actually come from a dataset where there are no FSRS users, they were all SM2 users.

1 Like

I can see the utility of this. If it doesn’t require large changes to implement, looks ok to me.

4 Likes

I dont really understand why this is a good idea

Wouldnt you want 2 separate presets? Whats the benefit of having 2 separate DR AND and mix the optimization?

Lets say you have special vocabulary you want to have higher DR. Maybe thats vocabulary you get exposed more because is more useful

At best it has no effect, and at worse, it is going to be worse

What makes that vocabulary special is going probably going to affect difficulty too, and therefore better have separated presets

(not against this change, just wanted to make sure i understand how this works)

  1. Too many presets are hard to manage.
  2. You can have smaller decks mixed in with bigger decks, and the small ones just benefit from having more reviews to generalise.
  3. FSRS-6 doesn’t show too much difference in performance when you have many/one preset (but that might be because most users learn one type of content, so not too sure about this last one).

Also, not sure your assumption always holds true. I might have some chapters that gets asked less in the exams but that doesn’t mean I’ll want a seperate preset for them. FYI I’m prepping for an exam that only takes place once in a year. The exam pattern has mostly nothing to do with the content. Our taxonomy section isn’t that important for example, but the content is on the same level as some other chapters.

Now that I think of it, I can make my DR a function of questions asked from this chapter per year value.

  1. im more concern with better use of the algorithm than management. Maybe presets made sense in the past, but post FSRS it seems to make more sense 1 deck 1 preset, as difficulties (what determines if you should separate or not) are usually going to change from deck to deck. It makes no sense for example optimize by subject, but to lump by difficulty. Presets made a lot of sense for the rest of the options but not for FSRS. Ill clone presets as much as needed for better algorithm
  2. This is also not true. The more you generalize, the worse the algorithm is going to perform. If im understanding correctly, it needs a minimum of reviews to make sense, but more reviews does not equal better algorithm. Just lumping cards with different difficulties to have “more reviews” is going to perform worse than allowing your reviews to grow organically
  3. Again, maybe im understanding all this wrong, but performance should be negligible. It takes resources to optimize a preset, not to use it. You only need to optimize sparsely

If you get ask much less irl about some content, that is going to affect how difficult those cards are. And that is bad for the algorithm. Mixing easy with hard the algorithm is going to draw something that catches both cases inside the same circle. But the circle is going to be bigger

There is no need to go crazy splitting content, but theres a few things that are easy to identify

1 Like

I have heard this one before, it can be put to test but I am sure it’ll fail.

If you don’t generalise, you get something that works well with the set of reviews that you’re optimising on but not for reviews you’ll do in the future.

I’m talking about the benchmarks, look them up.

If you’re taking my case, I didn’t appear for the entrance test every year in the past. So, the importance has nothing to do with my familiarity.

Well, I don’t see the point in arguing this. Some people for some reason might find this useful and some people might not. Both are valid.