Reinforcement Learning for Estimating Student Proficiency in Math Word Problems

Math Word Problems (MWPs) allow preparing students to apply mathematical skills in everyday life. Teaching MWPs is challenging because the mathematics difficulties of individual students vary across students within a class. In this paper, it is proposed a reinforcement learning algorithm based on Q-learning for estimating individual student proficiency according to the basic arithmetic operations: addition, subtraction, multiplication, and division. The proposal is analyzed by simulating interactions with hand-coded users but tested by simulating interactions with data-driven users. In addition, results are compared with those in a common Cognitive Diagnosis Model (CDM). Results show that our approach allows estimating the individual student proficiency in around 5 episodes. Thus, since reinforcement learning usually is applied to induce instructional policies in tutoring systems, this paper shows that it can be used for estimating student proficiency as well.