Ease-Hellish in FSRS

In a terrible turn of events, I seem to have achieved ease-hell in FSRS. To avoid any ambigiuity: It’s not the same ease-hell as in SM-2, where all hope is lost, but the convergence of pressing “Good” is so slow and insignficant, that it is by all means hell.

I don’t think this is easily reproducable, but maybe an interesting case study.
It’s my oldest japanese-learning deck, which actually followed my life from bachelor, over master to phd… So the review history is influenced full of different life-situations. Anyway, here is the interval graph from the visualizer, compared to the standard weights:
Mind the y-axis… And just to make sure: These intervals are absolutely too short for me, the cards don’t feel and are not that hard :smiley:

Share your parameters, please.

I did.

But here in copyable form :stuck_out_tongue:
0.4092, 0.4092, 0.4092, 8.3722, 5.7004, 2.5497, 0.4601, 0.0701, 1.2338, 0.4032, 0.536, 2.3644, 0.0447, 0.3702, 0.6782, 0.1824, 3.17

My bad, I didn’t notice at first.

0.4092, 0.4092, 0.4092, 8.3722, 5.7004, 2.5497, 0.4601, 0.0701, 1.2338, 0.4032, 0.536, 2.3644, 0.0447, 0.3702, 0.6782, 0.1824, 3.17

The parameter in bold determines how much difficulty reverts to the default value when you press “Good”. Seems like in your case it happens very slowly. I can’t think of any way to remedy this, other than increasing the value manually, but I highly doubt that you will get better metrics. Quite the opposite, log loss and RMSE will likely get worse.

2 Likes

Btw - I will simply use the default parameters for a few months now, those current parameters are spiralling out of control with the amount of work and backlog, making my review data less “real” with every week…

But it’s an interesting edge case, imho.

1 Like

In case you care about the inner workings of FSRS:
The parameter in bold (the difficulty multiplie Number 2, according to the visualizer) has almost no effect on my observed curves. I tried extreme values in both directions, and various values between 0 and 1.

The first three values have an extremly damping effect on everything, except if the first review is “Easy”. For that one case, it’s an cumulative effort of the other parameters, but again with very little influence by the parameter you bolded.

Yeah, difficulty in FSRS sucks. I don’t have much to say other than that.

Btw, that parameter is w7 here: The Algorithm · open-spaced-repetition/fsrs4anki Wiki · GitHub. Indexing starts from 0.

Those first 3 values are setting your initial Stability for cards that you grade Again, Hard, or Good the first time you see them. They are all getting the same value (which is unusual), and it’s very low.

But I don’t know what characteristics in your collection/behaviors in your review history are causing that optimization. Do you have any “bad” habits in your grading that you’ve left behind? Or did you used to have many, many learning steps?

Even if your collection is optimizable, you do have the option of using the default parameters for the time being. Depending on your Anki version, you can also optimize while ignoring older review history, if it is problematic.

On July 2nd, my first review for this deck will turn 10 years old according to the stats page - I reckon this deck experienced all bad habits imaginable :slight_smile: I did leave all those bad habits behind me though (the ones I know of, at least!) and followed the FSRS best practices so far. FWIW: Other, younger decks behave normally, both by feel and by curves.

I agree, default parameters are the way to go for now. Cherry-picking the perfect ignore-date is too much work, I think; At some point the ever growing backlog messed up my reviews due to being overworked and overwhelmed.

I just think this is a good example of solution space gone rogue. If it’s possible to pinpoint the behaviour to specific parameters being in unusual ranges, maybe users can be warned or advised to check the behaviour using the-under advertised-visualiser :smiley:

Ugh, yeah, a backlog makes everything harder! I find it makes it seem less oppressive if you make a plan for digging yourself out (what I do).

There is a setting to optimize using only recent review history. It might be a solution.