Recommending Mathematical Tasks Based on Reinforcement Learning and Item Response Theory

Recommending challenging and suitable exercises to students in an online learning environment is important, as it helps to stimulate their engagement and motivation. This requires considering their individual goals to improve learning efficiency on one side and on the other to provide tasks with an appropriate difficulty for the particular person. Apparently, this is not a trivial issue, and various approaches have been investigated in the areas of adaptive assessment and dynamic difficulty adjustment. Here, we present a solution for the domain of mathematics that rests on two pillars: Reinforcement Learning (RL) and Item Response Theory (IRT). Specifically, we investigated the effectiveness of two RL algorithms in recommending mathematical tasks to a sample of 125 first-year Bachelor’s students of computer science. Our recommendation was based on the Estimated Total Score (ETS) and item difficulty estimates derived from IRT. The results suggest that this method allowed for personalized and adaptive recommendations of items within the user-selected threshold while avoiding those with an already achieved target score. Experiments were performed on a real data set to demonstrate the potential of this approach in domains where task performance can be rigorously measured.