Use FSRS for spatial repetition?

[daposto]

Open Spaced Repetition

As far I understand FSRS is one of the best spatial repetition algorithms out there. Its claimed to be academically backed and it is open-source. ANKI is moving away from its own custom implementation. I would love if Hanzi-hero could use this or at least make it optional. Not sure if its worth trying to reinvent the wheel here, at least from the docs it reads like your making your own implementation.

[kevin]

Welcome!

This is something we looked into when we fixed up our SRS to be in-line with other SM-1 based SRS systems that most applications use. FSRS is still quite new, so we wanted to wait until it sees further adoption before making the jump ourselves. For example, it is not yet the defualt scheduler in Anki, as the lead developer himself notes that they want more time to make sure all kinks are ironed out.

[Axel]

I too would love this feature, it’s unfortunately keeping me from this otherwise awesome website

[daposto]

Current SRS is quite simple, other spatial systems have more options to indicate how well one remembers something. It allows me to schedule tasks with more granularity. Would be nice to have some more advanced options for indicating if I know or not know something.

Also would be great if I could have my ‘skipped’ be still put in the SRS but considered level expert.

[phil]

@daposto

other spatial systems have more options to indicate how well one remembers something

Would you happen to have an example? The current SRS indicators (e.g. Novice, Apprentice, etc.) are based off the interval until next review. I would be curious how other SRS systems signal to you how well you know something, beyond having the “stages” concept we have in HanziHero

[daposto]

Anki has defined intervals (“weights”) for how picking 1, 2, 3, or 4 for card’s level of difficulty. The difficulty becomes a parameter that a SRS can use to schedule cards more efficiently. In hanzihero’s terms, sometimes I know the meaning but not the sound, or the sound but not the meaning. Or maybe close to the real meaning. In those cases I like to tweak the tickets to be weighted.

[kaysik]

@daposto

Anki has defined intervals (“weights”) for how picking 1, 2, 3, or 4 for card’s level of difficulty

FYI - research done by the FSRS team has found that those using all 4 buttons (again, hard, good, easy) actually do worse than those only using 2 buttons (again + good). It’s currently a bit unclear exactly why, but the current hypothesis is people abuse “hard” when they shouldn’t and aren’t consistent with good vs easy. People are more emotional than we like to claim - so hitting “hard” today because “I didn’t really get it wrong” is very easy but screws up the SRS scheduling because you actually didn’t know it fully, you just don’t want to feel bad. Or seeing “1 month” as an interval is scary so hit hard when it wasn’t needed etc. SRS systems work best when you’re 100% consistent in how you answer. When it’s binary choice it’s much easier to be consistent and honestly about if you knew it or not.

The simplicity of HH’s just “type it in, hit enter and you’re either wrong/right” works great. I’d hate to lose that and have it ask you to rank things after every answers. There is probably no benefit to the extra choices and it would slow down study sessions considerably.

[phil]

@daposto

difficulty becomes a parameter that a SRS can use to schedule cards more efficiently

Ah, I understand now. I thought we were talking about labels, not actually influencing the algorithm itself – got it. I don’t see us offering more options in influencing the SRS algorithm at the moment, it’s hard to see how it’d fit into the right/wrong of typing

@kaysik

When it’s binary choice it’s much easier to be consistent and honestly about if you knew it or not.

Exactly, keeping it simple and reducing the number of decisions drives our design. We had another thread which talked about this which provides some interesting background:

Way to make reviews take up less time

I am reminded of the whole “low-key anki” series, which you can read through the beginning here if you’re interested. There are two key insights I would extract from it. We can’t really judge how well we know something Not to mention the ever-present possibility that the only reason you found a card easy to answer in a specific moment was due to the same sort of memory fluke that I talked about in the previous section. In fact, ja-dark actually cited studies that found that people’s subjectiv…

[KhalilK97]

You guys might want to give FSRS another look. 5.0 released recently and they’ve got a lot of data from Anki users backing it up: GitHub - open-spaced-repetition/srs-benchmark: A benchmark for spaced repetition schedulers/algorithms

Having that big decrease in review load would sure be nice… (Or perhaps a big increase in daily lesson load? )