Why did these convergences fail? Here are some thoughts.

[L-M-Sherlock]

Here is a case:

The parameters are:

[0.3059, 1.7308, 13.2325, 28.9381, 7.0016, 0.383, 2.2044, 0.001, 1.5403, 0.4922, 1.0743, 1.8826, 0.1139, 0.3747, 2.5531, 0.2114, 2.4432, 0.5819, 0.9792]

Here, the cost refers to the expected time required for a card to transition from its current memory state to the target memory state.

The red region implies that the cost should be higher than 1000.

According to my another simulation, it will take 293011 seconds to reach the target memory state if the current stability is 0.1 and difficulty is 10.

[L-M-Sherlock]

I write a new notebook to analyze the convergence: https://github.com/open-spaced-repetition/SSP-MMC-FSRS/blob/main/convergence_analysis.ipynb

After analyzing the distribution of parameters, I noticed some significant differences:

It means unconverged users' memory stability increases more slowly than convergded users.