I have been using FSRS for around 2 years now, and find it fantastic in predicting my desired retention and helping me learn cards. I can always recall almost exactly 90% of all the cards at all times.
However, I don’t need 90% of the cards memorized at all times. I just want to spend the least possible time learning everything. If I only know 60%, thats fine - the day before my exam I can just do a filtered deck and I will know 100%.
Therefore, I would like to have a mode in Anki/FSRS where instead of choosing perfect intervals for your desired retention, it would optimize interval lengths for least time spent to get cards into your long-term memory.
This would likely be lower at the beginning, where forgetting a card is not so bad and you can just relearn it. Once a card is very stable, even 99% DR would give long intervals. When forgetting those cards, you spend a lot of time getting the card back up to the very long intervals. Thus the OI algorithm would shorten very long intervals, and enlongen the shorter intervals.
Is there any fork of anki or fsrs implementation for this?
Now you’ve reached the classical education system. You study for five years, and then you have a general exam covering all the subjects you’ve studied.
So, I think it’s a good strategy to study the cards for up to 21 days, and then repeat only as needed, at longer intervals. But this violates the logic of Anki’s FSRS algorithm. It’s simply its own method.
Another approach, which Anki also recommends, is to avoid studying everything—that is, to create cards for everything, only for the most important and the hardest to remember. This method allows you to reduce the number of cards by a factor of 10, and instead of having 20,000 cards, you’ll have just 2,000 that are easier to review a month before the annual general exam.
That’s exactly what L.M.Sherlock (Jarrett Ye), the creator of FSRS, proposed in his 2022 research paper (that scheduling algorithm is called SSP-MMC).
However, the subsequent experiments have not been able to conclusively prove that SSP-MMC would outperform a fixed DR algorithm.
Consequently, Jarrett has not created any ready-to-use implementation of the SSP-MMC algorithm. However, I think that one user has created a custom scheduling script for the same, not sure if they are on this Forum.
When I see a description like this “Experimental results have shown a 12.6% performance improvement over the state-of-the-art methods. The scheduler has been successfully deployed in the online language-learning app MaiMemo to help millions of students.” I always ask myself, why not 50%, 100%.
It seems like a tenth of a percent saves you time. But mathematical data is just an average. For some, an algorithm will yield 50%, for others, 0%. Everyone should test any method for themselves and use what works for them. And good teachers say there’s no unique learning method for everyone, because some people are interested in a topic and absorb it quickly, while others are forced and barely cope. Sometimes it’s not about intervals. Sometimes you need to find a special approach or method that you enjoy, and using this method, you’ll see rapid progress.
And statistics is a lie, and they use it to deceive people (I studied statistics for a couple of years, held a certain position, so I know how it works, and there’s a book called “How to Lie with Statistics” – so please don’t be offended if I occasionally criticize certain methods. I know the research was paid for, and I need to show the results.) .
.