[Question] Explain how the optimizer calculates retention that minimizes review times about fsrs-optimizer HOT 12 CLOSED

By the way, I plan to use a brand new way to find the optimal retention in FSRS Optimizer. It has been implemented in https://github.com/open-spaced-repetition/fsrs-rs

The main idea of the new method is to simulate the review process of user with different retention and select the retention which maximize the estimated total knowledge.

from fsrs-optimizer.

The forgotten cards' ratings must be again, so we don't need to sample for them. The code only sample rating fromhard, good and easy for those recalled cards.

Here I have estimated these probabilities from the user's actual review history.

from fsrs-optimizer.

Please see my paper. The mechanism of this part has been described in Section 4 OPTIMAL SCHEDULING:

source: www.maimemo.com/paper/

from fsrs-optimizer.

The main idea of the new method is to simulate the review process of user with different retention and select the retention which maximize the estimated total knowledge.

Can you explain how that one works, then? The way I see it, it will just output the max. retention every time, since that's what maximizes total knowledge. Unless you mean total knowledge acquired per unit of time.

from fsrs-optimizer.

In the simulation, the time to learn per day is fixed. If the retention is too high, the user will have no time to learn new cards.

For the details of simulation, please see this pseudo code:

from fsrs-optimizer.

That's not very easy for me to understand, but thank you.
Unrelated, but have you compared the speeds of the Rust-based optimizer and Python-based optimizer? I would assume that Rust version is faster.

from fsrs-optimizer.

I haven't compared the speeds. Because Rust-based optimizer hasn't implemented the splits of dataset.

from fsrs-optimizer.

Also, how do you know whether the new method is better?
With the algorithm itself, finding out whether a change is good or bad is very straightforward - just run both on the same dataset and check the RMSE. We have "ground truth" - the actual repetition history. But how do you assess which method of finding optimal retention is better?

from fsrs-optimizer.

The old method didn't support decimal difficulty, so it's imprecise. And it also assumed the user only press good and again.

from fsrs-optimizer.

So I was looking at the code, and this raises 2 questions:

1. If I understand it correctly, does it mean that in the simulator "Again" can only happen during the first review, and cannot happen during later reviews? review_rating_prob only has 3 values.
2. Wouldn't it be better to estimate these probabilities from the user's actual review history?

from fsrs-optimizer.

Thank you. One more question: does the simulator use the same value of answer time (recall_cost) for Hard, Good and Easy? Wouldn't it be more precise to use 3 different values for three passing grades?

from fsrs-optimizer.

The current simulator uses the same value. You can open an issue for this feature request.

from fsrs-optimizer.