Description
In Spring 2018, students in UC Berkeley’s introduction to computing for non-majors course CS10: The Beauty and Joy of Computing tested a prototype Intelligent Tutoring System known as AutoQuiz. It was customized to the assessment material in CS10, and was designed to model user knowledge, then use that information to adapt to individual students and serve multiple-choice questions that match in difficulty with the student capabilities to help students prepare for high-stakes assessments.
In this paper, we present a complete overhaul of the artificial intelligence algorithms that power AutoQuiz in order to increase its ability to serve students. We compare the previous “adapted DKT model” approach against a new deep-reinforcement-learning-based system, which we call Deep Knowledge Reinforcer (DKR). While the previous adapted DKT model only attempts to track student knowledge, the Deep Knowledge Reinforcer model attempts to both model a student’s current knowledge and determine how to increase that knowledge most effectively. In order to match students’ knowledge more accurately, we enhanced the questions by adding different formats, which can give expanded insight into student mental models and misconceptions. These improvements were tested during the Fall 2018 CS10 course offering with over 100 students creating anonymized accounts.
We show that student use of this iteration of AutoQuiz is correlated with a marked improvement in exam performance, and thus provides tentative evidence for the claim that a reinforcement-learning-based system can effectively work to teach students. A future randomized controlled experiment could be used to demonstrate a causal link.