Hi all,
A release candidate is now ready to test. Please let us know how you go.
(it may be renamed 24.11 when the final stable release is ready)
Download: Releases · ankitects/anki · GitHub
Previous discussion: Anki 24.10 Beta
- Don’t exclude suspended cards when updating all FSRS params.
If I understand correctly, the underlined text should be missing.
The change in question was about applying memory state updates to cards that happen to be suspended - it’s not about the cards that are used to calculate FSRS parameters.
I don’t know how it should work. the checkbox does not remember the status.
The entry just before that shows an interval of 1d which made it go to the review state. The next lapse happened when the card was a review card. Then it became a learn card again. It’s normal behaviour I think.
(I hope we have no problems here, I think 24.11 as version name sounds weird)
1 Like
Is it normal that you can reduce the RMSE so much when manually editing the weigh?
Weigh Optimize
Manually changed w0-w3 based on the “Evalute” readings
The same thing was noticed in FSRS 4.5, the improvement was negligible.
1 Like
I am not sure if you should here. See log loss is actually increasing.
I overdid it. Here’s a simpler change. If you spend a little more time on this, the result will be even better.
1 Like
I have seen somewhere that some user found out that using the standalone FSRS optimizer gives better parameters (presumably because the optimizer has more time 
). I think Sherlock should have a look at this. I dont think the Anki optimizer gives the best parameters.
I’ve been able to see some improvements in metrics recently after meddling with my w[6], mostly through decreasing it. Described it in discord. You might try that as well. But yes, 1 percentage drop when RMSE is 5% sounds a lot.
That was me and that’s still the case. It’s not always better, but there is a chance. There is still room for improvement even there, as you can optimize with a random seed (see next paragraph), which can sometimes give better parameters after a few tries, with the downside that optimizing doesn’t give the same result every time you do it.
A seed is a number that influences the result. With the same seed and the same set of reviews, you get the same result. With a different seed, you might get a different result, but there is no way to know beforehand if the result will be better and there is no “optimal seed” for all circumstances.
I’m currently performing some tests to see if I can find an improvement to how it currently works the apps, without getting a different result every time you click and without doing more computations. I have an idea, but it might be flawed. Over a longer time period it shouldn’t make a difference anyway.
So I want to emphasize: based on what I know, there is no reason to believe that the implementation that’s used in the app is worse. It’s just different and there should be a 50/50 chance that the Python optimizer gives a better result with the same set of reviews. So if the Python optimizer is better today, it might be worse next week.
1 Like
It’s a known issue:
FSRS-5 is the Python version. In 55.4% cases, it’s better than FSRS-rs (the Anki optimizer), but worse in 44.6% cases.
2 Likes
As another user said, it’s not the amount of time that affected the result, but the seed (well, plus the fact that the Rust version is consistently slightly worse than the Python version, as Jarrett pointed out above).
Changing the seed has a 50/50 chance of improving/worsening the result. However, I proposed a method that, combined with keeping the old parameters if the new ones are worse, could improve optimization at no cost.
Idk if Jarrett will test it through, and I can’t test it on my own.
(also, this is getting kinda off topic. I described my idea here: Anticipated Features in 24.10 - #38 by Expertium)
1 Like
When you right-click the mouse, the interface stops responding until you make a left click.
I tested it here and observed the same behavior. When I right-click the mouse, the interface indeed stops responding until I make a left-click. I believe this happens because the system is creating an empty dropdown menu. Since I use an addon that adds an extra option (attached image), the menu appears normally. However, when I open it without the addon, the menu doesn’t show up, and the system appears to be stuck, just as you described.
1 Like
my too! 
So funny bug
Yes, that was indeed what I meant when I proposed to tie the seed to the review count in the original thread. I’m already running some tests on a small dataset, but I have no conclusive results yet and with my dataset it will probably never be conclusive anyway, but I was curious it would point in some direction.
I see the same behavior in 24.06, but the context menu is never empty. So the change appears to be that the Copy option is gone from the context menu?
Nothing appears.
1 Like
@dae, could you wait for this patch of FSRS?
I have figured out how to solve it. And it will make FSRS-rs as accurate as the torch version.
8 Likes