BieniekAlexander
commented
1 year ago I think the questions you've asked here are missing the simplicity of the algorithm. All that Ebisu does is model the half-life of the recall of a fact, stating that a user is expected to recall a study fact in half_life time with 50% probability. (I might be mincing statistical phrasing, but this is the gist). You can use Ebisu to update models according to a user's response to a fact in the future, and the new model is meant to more accurately represent how long that half life is.
Ebisu doesn't say anything about when you should review your fact, whether the card is new or not, etc. etc. As a user of Ebisu, it's sort of up to you to understand how strong(?) the priors of the user's memory are, and how quickly your user will forget the term. I do think that this project could do with some more guidance on how to set a model's initial parameters, but as these parameters are modeling a human's ability to remember something, it's basically impossible to say what the perfect parameters are. Given that a term would be a new fact, the authors seem to suggest that, to model a user's remembering a random word, it's okay to start with a model represented with (alpha=2.0, beta=2.0, half_life=24.0) (where half_life is in hours).
Does this new algorithm remove all reviewed cards and resets them as new cards? Does it assign any kind of estimate for reviewed cards or can it only be used for a new deck of cards? Like for example starting a new deck from scratch?