An ACT-R Model of Memory Applied to Finding the Optimal Schedule of Practice

When a person confronts the task of memorizing a collection of facts some questions have to be answered about how to optimize their learning. Perhaps the most important of these questions is how the practices for each item should be scheduled. The most basic suggestion from psychology is that the practices should be spaced as widely as possible. Since Bahrick (1979) it has been clear that wide spacing intervals result in better performance after a long delay. However, while wide spacing is likely to help, it may also be important to consider item differences when making practice scheduling decisions. Memory items (such as the Japanese-English pairedassociates we have used in this research) are not equally difficult for each subject to learn. In postexperimental interviews subjects often mention that some items were easier to remember because they could find some mediator to connect the cue to the response. Further, the data shows that some items are on average easier for all subjects. Perhaps the cue (the Japanese word) for these items is fairly close to some common English word that can be used to mediate between the cue and response. It is useful to consider how we might optimize learning by attending to this variability in item difficulty. For instance, if we knew an item was easier we could schedule it less often, while if it was more difficult we could schedule it more often. This may have two benefits. First, if we did not use this procedure and instead gave the items equal practice we might expect that some items would be overlearned (the easier items) and some items would be learned poorly (the harder items). This uneven learning might be undesirable. Second, by saving practices on easy items and then distributing them to hard items it might be possible to achieve gains in performance. To attend to this variability requires a model of memory that does two things. First, it needs to describe accurately how practice in general leads to recall performance by characterizing memory effects due to the frequency, recency, and spacing of practices. Second, to optimize performance for individual items, the model needs to be able to characterize both initial average difficulty for each item (estimated by fitting data from prior experiments, from now on referred to as item difficulty) and the difficulty particular subjects have with each item (which will need to be estimated based on performance with the item over the course of each subject’s learning, from now on referred to as subject/item difficulty).