Intelligent Tutoring Systems (ITS) have gained increased attention, especially during the shift to online learning driven by the COVID-19 pandemic. However, maintaining student engagement remains a significant challenge, as students often disengage by skipping problems or excessively relying on hints, which hinders their learning progress. To address this issue, we analyzed MathSpring student log data, revealing that only about 10% of student interactions led to a successful transition from an initially unsolved problem to solving the subsequent one. This paper introduces a novel reinforcement learning (RL) model integrated into the MathSpring tutoring system to combat disengagement and optimize problem recommendations. Using a dataset of over 300,000 student interactions, we trained a Q-learning-based model to personalize problem assignments based on each student’s skill level and engagement tendencies. The model dynamically adapts recommendations in real-time by leveraging feedback from student interactions, ensuring that students encounter problems of appropriate difficulty. This personalized approach aims to promote mastery learning by maintaining optimal engagement and providing targeted feedback aligned with individual comprehension levels. Additionally, the Q-table generated by the RL model offers a valuable data mining tool for identifying effective teaching interventions, allowing researchers and educators to analyze which instructional strategies most consistently lead to positive learning outcomes. We present the design, implementation, and initial evaluation of the RL model within MathSpring and discuss its potential benefits for improving student engagement. Plans for future evaluations to assess long-term effectiveness and learning outcomes are also outlined.

Keywords: Learning technologies, Students engagement, Reinforcement learning, Educational data mining, Math education, personalized learning.