Behavioural Cloning of Teachers for Automatic Homework Selection

Abstract

We describe a machine-learning system for supporting teachers through the selection of homework assignments. Our system uses behavioural cloning of teacher activity to generate personalised homework assignments for students. Classroom use is then supported through additional mechanisms to combine these predictions into group assignments. We train and evaluate our system against 50,065 homework assignments collected over two years by the Isaac Physics platform. We use baseline policies incorporating expert curriculum knowledge for evaluation and find that our technique improves on the strongest baseline policy by 18.5% in Year 1 and by 13.3% in Year 2.