[BUG] [FSRS4Anki v1.6.1 Optimizer] Optimized Default Difficulty is too low on large revlog about fsrs4anki

[L-M-Sherlock]

But the default difficulty deceased monotonically in your case. I will test your collection tomorrow.

[kuroahna]

Hm actually, you might be right. A large majority of my cards have ~250% ease

And looking at the interval history if I keep pressing Good:

first rating: 3
rating history: 3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3
interval history: 0,4,10,25,56,119,240,460,845,1493,2551,4227,6812,10707,16451,24757
difficulty history: 0,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2,1.2

The change between each interval is:

0 -> 4 = N/A
4 -> 10 = 10 / 4 = 2.5 = 250%
10 -> 25 = 25 / 10 = 2.5 = 250%
25 -> 56 = 56/ 25 = 2.24 = 224%
56 -> 119 = 119 / 56 = 2.125 = 212.5%
119 -> 240 = 240 / 119 = 2.017 = 201.7%
...

This seems correct since most of my cards have 250% and I have about ~88% retention rate for my young cards. The FSRS4anki algorithm starts with 250% change for the interval and then it slowly decreases the change in interval as I keep pressing good.

In version FSRS4Anki v1.4.3 Optimizer however, I got the results:

let defaultDifficulty = 3.835;
let defaultStability = 3.2191;
let difficultyDecay = -0.9131;
let stabilityDecay = -0.0872;
let retrievabilityFactor = 1.2571;
let increaseFactor = 3.2025;
let lapsesBase = -0.0493;

stability, difficulty, lapses
     3.22        3.83       0
     5.77        3.84       0
    10.62        3.84       0
    19.06        3.83       0
    32.91        3.83       0
    55.86        3.83       0
    93.04        3.83       0
   152.10        3.83       0
   244.58        3.83       0
   387.58        3.83       0
   605.15        3.83       0
   931.47        3.83       0
  1415.08        3.83       0
  2123.76        3.83       0
  3150.51        3.83       0
0,3,6,11,19,33,56,93,152,245,388,605,931,1415,2124,3151

The change between each interval here is:

0 -> 3 = N/A
3 -> 6 = 6 / 3 = 2 = 200%
6 -> 11 = 11 /  6 = 1.83 = 183%
11 -> 19 = 19 / 11 = 1.73 = 173%
19 -> 33 = 33 / 19 = 1.74 = 174%
33 -> 56 = 56 / 33 = 1.70 = 170%
...

With these parameters, it starts with a change in interval of 200%, not 250%. It does seem like these new parameters are correct. I guess the larger learning rate allowed it to converge faster.