Read Full Report
This project was developed with the help of Winter Clark as the Educational Researcher and with the mentorship of Dr. Vincent Aleven, Jonathan Sewall, Kenneth Hirsh, and Nicholas Greenberg.
After several years of development, Ella’s musical curriculum has grown into over a thousand exercises ranging from intervals to long stretches of music. When students use the practice mode to test their skills against the exercises, they are exposed to random questions sequentially and indefinitely. This presented several recurring problems:
- When there is too much content, what design decisions lower the cognitive load when supporting the student in selecting what to do next?
- When the experience is of indefinite length, what design decisions prevent vocal injury while supporting the practice?
- When there is a need to communicate proficiency, what is the correct framework to do so, and how can we support the student’s self-awareness?
The concept was to create a mechanism that could select the next best activity with an acceptable degree of certainty, just like a tutor or experienced advisor. Instead of implementing a whole system at once, we narrowed an experiment aiming for high impact and interpretable insights on the next steps: what is the best question a student can be exposed to at any given point that will further their musical skill?
We employed artificial intelligence, namely Bayesian Knowledge Tracing (BKT), in a tagged set of student tasks. This technique allows us to estimate the next step in the shortest path to mastery, thus efficiently using the student’s time and effort.
The result of this study was a practice mode that selected exercises the students would benefit from with better-than-random accuracy. Additionally, the technique estimated when it was best to exit the practice early due to high proficiency while stopping long sessions with no improvement using the mastery data.