Spaced Repetition Knobs and Tweaks

Spaced Repetition Knobs and Tweaks

Postby Vinde » Sat Mar 16, 2019 2:36 am

Hi!

I'm Learning Kanji for the long term using spaced repetition, and since I've just started this deck I thought it's a good time to know what SRS setting you guys use/recommend.

Just yesterday I read someone's recommendation that they use Fibonacci intervals for SRS (1, 2, 3, 5, 8 ...). That, iirc, would mean that we use a "Correct Interval Factor = φ = 1.618..." and a "Wrong Interval Factor = 1/φ = 0.618..."? Not sure what to set for "Strong", or if "Early" should also be equal to 1.618.. to be accurate to Fibonacci. (I think FD rounds the numbers to one decimal place, which is not an issue as 1.6 is a good approximation of φ)

I've never used Anki enough to know what numbers does Anki use.
What settings do you guys recommend? Is there any research into what numbers are ideal, I assume that would be a bell curve, but still? So far, I'd been using the default values which are pretty good anyways lol

Looking forward to a good discussion!
Vinde
 
Posts: 14
Joined: Thu Mar 14, 2019 5:42 am

Re: Spaced Repetition Knobs and Tweaks

Postby Ernie » Sun Mar 17, 2019 8:21 am

Hi Vinde,

I think Anki's formulas are based off of Supermemo's, which are very complex. I didn't want to go down that route because I wanted something that people can relatively easily understand and tweak. I came up with my default intervals based on what I saw others doing (including Supermemo) and added some conservativeness to it. From my readings back then, I was, and still am, under the impression that little differences in calculations across apps do not affect results much. One goal of SR is to show the card again just as you about to forget it. Supermemo, I think, will dynamically change next intervals to target say a 90% correct rate. I don't have anything auto-adjusting interval calcs like that, but you can view "Correct % by Interval (SR)" in the Stats screen to get a sense of where you are at.

If you feel you are getting tested too often, increase interval factors, increase initial interval, swipe up more often, etc. I use the default factors myself, but I forget things easily, so others may find them a bit too conservative. I change Wrong Interval Factor to 0.7 though so wrong answers aren't so detrimental. I set "new to active after" to "1 correct", so I can manually control when I want each card to become active. I added a new "Postpone / Next Review" "study button" to the app which you can add to the bottom of the study screen to make it a little easier to make overrides to default behavior.

I have an "Early correct factor" because I feel intervals should rise more quickly (factor-wise) at the beginning. Strong is for swipe up, so if you know it well, delay it longer.

I don't really have much of a conversation available as to what is the best, other than what I wrote above. Personally, I'd wouldn't worry too much about it and just go with what feels right after using the app a while. As you get more cards in your deck, it's possible that you're going to start getting overwhelmed with due cards. At this point, you'll start to think more about how to get intervals up more quickly (using options I mentioned above), but you of course don't want to get them up too quickly that you are forgetting a lot. Since each person is different and each subject as different in difficulty, I doubt there something that is "best". This is maybe where the "dynamic interval calc" that I mentioned above might be nice (that would auto raise/lower factors depending on how you are doing), but again, I didn't want the complications of something like that, both for me and the user.
Ernie
Site Admin
 
Posts: 5009
Joined: Fri Oct 01, 2010 6:12 pm
Location: San Francisco / Taiwan

Re: Spaced Repetition Knobs and Tweaks

Postby kraemder » Sun Mar 17, 2019 2:46 pm

I would encourage experimentation. A lot of people will just use the default settings because they don't want to worry about it but later on you may find that a very different setting works really well for you. If you're cramming for a test for example you would probably want to do lower intervals as it's critical to make sure you know -everything- to get the best score. But if you're just learning vocabulary for the long term... having much larger intervals may work better to keep it from getting too boring from seeing the same material over and over and over and over. And always maintaining that near perfect recall won't be as important.

I think Anki's formula is based on an older super memo routine from years ago and super memo has changed their formula a lot since then. Or so I hear. So I don't know that I would consider Anki's formula really advanced. And I doubt that really fine tuning a formula I going to make a huge difference in your studying since every vocabulary or kanji really has its own unique learning curve. You say you're new to this so you may have heard bout leeches. One thing I like about FCD over Anki is the way you can handle leeches. FCD doesn't flag them for you which Anki does but with FCD you can manually just use 'next review' and push it out for a month or several months - because right now this word is -not- sticking and you're just not ready to learn it but later (whenever that is) you should be. With Anki, it will suspend it. Permanently so that using the desktop app you have to go in and find it and reactivate it manually. I hated that and consequently I end up turning off the whole leech function entirely.

Anyway, I used the default Anki settings for years and then questioned it. I don't understand the way Anki lets users tweak the defaults etc., so I went onto Google and found some posts from people who tweaked the settings to greatly increase the interval settings and were really happy with the results. I copied them and I was really happy too and wish I had done this earlier. I did this for Japanese which I had been studying for years. I tried doing the same thing for Korean when I started studying it Thanksgiving late last year and I was forgetting everything and maybe that wasn't such a big deal but I lowered the interval settings a lot until I've studied Korean more.

Good luck. Judging from a quick look at Fibonacci number sequences I can see why someone would think they would make a good SRS stepping system. But I don't think it's all that different from just say multiplying by 1.6 or 2.0 or whatever. I don't know of any SRS apps using it per se.
kraemder
 
Posts: 445
Joined: Sun Oct 31, 2010 1:08 pm

Re: Spaced Repetition Knobs and Tweaks

Postby Ernie » Mon Mar 18, 2019 5:09 am

Thanks Kraemder on your feedback and views!
Ernie
Site Admin
 
Posts: 5009
Joined: Fri Oct 01, 2010 6:12 pm
Location: San Francisco / Taiwan

Re: Spaced Repetition Knobs and Tweaks

Postby Vinde » Tue Mar 19, 2019 11:26 pm

Thanks for dropping by to share your insights!

@Ernie, hearing from you that you've increased the Wrong Interval Factor (0.5 - 0.7) gives me the confidence to do the same :D. Placebo or not, now it doesn't feel as bad when I can't recall an answer. That makes the learning process feel less stressful and more positive!

As a side ramble: I was just thinking about the whole human memory model etc. and I was wondering about a hypothetical scenario.

Say we have an immortal human being, Alice, who's guaranteed to live forever (no accidents, nothing bad happens).
Now, I think we can agree that given enough time (call it T0) Alice will forget her mother tongue if she stops using/hearing it.
I have a hunch that T0 will approach a fixed value. What I mean is that it doesn't matter if she has used her mother tongue for a trillion years or a centillion years, her T0 will always be around, say, 3000 years.

If you agree that T0 approaches a constant value, then you will agree that, although not practically (since humans don't live forever) but theoretically, Correct Interval Factor (CIF) should asymptote to 1.

Now, I'm not saying that we change our algo :P, the decrease in CIF is non-significant over a few years (?). But I think that it's pretty cool that in a way FCD already approximates that curve by decreasing the interval factor from Early Correct Factor to CIF.

@kraemder: I've been experimenting now that my 'active' list keeps growing. I actually did not know about leeches! But I see what you mean, I'll look into the 'next review' function. FCD also feels 'nice' as it doesn't seem to get in the way of your learning.
Vinde
 
Posts: 14
Joined: Thu Mar 14, 2019 5:42 am

Return to Comments, Questions and Feedback