Ask questionsAnki Implementation of Ebisu algorithm
I just went quickly through your note, but it seems like an excellent and math-based approach. Great Job! I was wondering whether I could implement your scheduler in the Anki app. How much would you quantify the effectiveness? Correct me if I am wrong, but the primary benefit that I see is that you can handle better over-studying and under-studying. In case one follows the card schedule diligently, how much time in review the cards (without reducing recall) can be saved?
Have a nice day
Answer questions TommasoBendinelli
Thank you for the great answer and congrats again for this super repo. The code of the paper is available here https://github.com/rddy/deeptutor, although it is just a large Jupiter notebook.
I agree with you that simulating students just provided blurry confidence in algorithm performances, but apart from a real experiment, it is the best evaluation method we can get. Also, I think that although all the factors that you mention are important for improving the learning experience, I also believe that a "great" SRS algorithm can make a difference, especially for "mature" cards (i.e., cards with long-expected half-life).
My end goal would be to create an SRS algorithm based on Reinforcement Learning (it sounds a bit fancy), not only to predict the recall but also to automatically scheduling cards review. This algorithm would be, correct me if I am wrong, a bit different from Ebisu, where the recall threshold for performing the review is fixed. Ideally, the algorithm should find the optimal trade-off between maximizing recall probability and reducing the number of reviews.
Currently, I am exploring the field, although the literature is a bit scarce. Besides your repo and the mentioned paper, I am looking at the Duolingo half-life regression algorithm and this blog: https://papousek.github.io/modeling-prior-knowlegde-using-duolingo-data-set.html. Probably, you are already aware of these resources.
I have in mind to start coding in the next days by creating a reliable benchmark for evaluating the different algorithms and approaches (definitely Ebisu), similar to what the mentioned paper has done.
I would be delighted if you want to work jointly on this idea. If you are up to, we can even arrange a call to discuss the details.
Related questionsNo questions were found.