This paper attempts to produce a better scheduler for a Spaced repetition memory system by combining a model-based (per-student, per-item) estimator of student memory state, with a more holistic planning model which attempts to take students’ time constraints and test date into account.
They compare a simple Leitner model with a model-based recall estimator, and compares two planning approaches in the latter case: the first is a simple threshold based approach (like in Mozer, M. C., & Lindsey, R. V. (2016). Predicting and Improving Memory Retention: Psychological Theory Matters in the Big Data Era. In M. N. Jones (Ed.), Big data in cognitive science (pp. 34–64).); the second is a holistic planning optimization. In a study of students learning Kanji, the former does better. This is awfully convenient, since it’s simpler to implement and more flexible!
Another key difference from Mozer and Lindsey (2016) is that they infer per-user, per-item parameters—so this is a much higher-dimensional model than the DASH model, which infers a single per-user and a single per-item parameter.
Q. What’s the primary conceptual difference in scheduling approach between this paper and Mozer/Lindsey (2016)?
A. They take the students’ time constraints and test schedule into account.
Q. Compare the dimensionality of this paper’s prediction model with that of Mozer/Lindsey (2016).
A. This paper learns per-user/per-item parameters, while M/L learns separate “difficulty” and “ability” parameters—roughly square root as many.
Q. What was the key result of this paper’s human study?
A. A threshold-based (“myopic”) planner did better than a more complex holistic planner (and also better than a simple Leitner model).