Feat/FSRS-5

[L-M-Sherlock]

candidate:

• Feat/consider short term schedule [WIP] #81
• Feat/exp initial difficulty [WIP] #110

I plan to release FSRS-5 in June.

[Expertium]

I plan to release FSRS-5 in June.

I think that instead of aiming for "release FSRS-5 in [month]", we should aim for "release FSRS-5 when RMSE decreases by at least 10% relative to FSRS-4.5"

[L-M-Sherlock]

I think that instead of aiming for "release FSRS-5 in [month]", we should aim for "release FSRS-5 when RMSE decreases by at least 10% relative to FSRS-4.5"

I'm afraid FSRS-5 will be never released if we need to achieve that target.

• FSRS-4.5 was released 5 months ago.
• FSFSv4 was released 10 months ago.
• FSFSv3 was released 20 months ago.

I think it's a good time to release FSRS-5.

[Expertium]

With all due respect, I don't understand why you feel the need to release a new version if it's barely an improvement compared to the previous version. I firmly believe that we should aim for at least 10%. More would be better, of course. To me, 10% seems like the bare minimum; below that point, it's just not worth bothering with releasing a new version.

[Expertium]

@user1823 out of curiosity, what's your opinion?

[user1823]

As the algorithm improves, the scope for further improvement decreases. So, aiming for a 10% improvement every time is not feasible.

If the algorithm is going to consider short-term reviews now, I believe that this is a significant change and the update deserves to be called v5.

[ishiko732]

I noticed that there is a change in the DEFAULT_PARAMETER length. How should migrate from FSRSv4.5 to FSRSv5? Do just need to add the default parameters w[17] and w[18]?

[user1823]

if by October we won't find a way to decrease RMSE by >=10%, we will release FSRS-5 in October, otherwise we will release it earlier?

Makes sense. I agree that we don't need to rush through the release. Having longer intervals between the releases means that

• we can test more ideas and select the more effective ones
• it is less confusing for the users because incompatible (whether partially or fully) versions of desktop and mobile apps cause a lot of confusion

[L-M-Sherlock]

if by October we won't find a way to decrease RMSE by >=10%, we will release FSRS-5 in October, otherwise we will release it earlier?

OK.

I have some criteria for the ideas:

1. If the idea introduce extra parameters, the improvement should be greater than 2% per param.
2. It should respect the current definition and range of stability and difficulty.
3. It should be compatible with the old parameters.

[Expertium]

1. If the idea introduce extra parameters, the improvement should be greater than 2% per param.

That's gonna be hard.

2. It should respect the current definition and range of stability and difficulty.

Does that mean that my new definition of D (based on R) is automatically disqualified?

3. It should be compatible with the old parameters.

What do you mean exactly?

[L-M-Sherlock]

Does that mean that my new definition of D (based on R) is automatically disqualified?

If your new definition of D is out of the range 1~10, it would be rejected.

What do you mean exactly?

The user could use the new parameters without rescheduling all their cards.

[Expertium]

The user could use the new parameters without rescheduling all their cards.

This one will always be satisfied, no? I mean, the user will have to click "Optimize" to obtain new parameters, but won't have to reschedule. How would it be different?

[Expertium]

forget_rating_offest=-0.1

Typo, it should be forget_rating_offset. Btw, what does it do?

[L-M-Sherlock]

Typo, it should be forget_rating_offset. Btw, what does it do?

Sometime the user may press again during the relearning step, it will reduce the stability in FSRS-5. The forget_rating_offset is used to simulate it.

[Expertium]

Why not just estimate 4 values of answer button probabilities, but for learning and re-learning steps?

[L-M-Sherlock]

Why not just estimate 4 values of answer button probabilities, but for learning and re-learning steps?

It's computing-expensive and hard to model.

[Expertium]

Is it? It's just another numpy array (and whatever equivalent of a numpy array is used in Rust), similar to first_rating_prob. maybe you misunderstood me, I meant "Make an array with 4 values, like first_rating_prob, but for probabilities of pressing Again/Hard/Goode/Easy in (re)learning"

[L-M-Sherlock]

Is it? It's just another numpy array (and whatever equivalent of a numpy array is used in Rust), similar to first_rating_prob. maybe you misunderstood me, I meant "Make an array with 4 values, like first_rating_prob, but for probabilities of pressing Again/Hard/Goode/Easy in (re)learning"

The user may not only press once. The short-term rating sequence would be 1,2,2,3, 1,3, 1,4, 1,1,3,1,3 and so on.