Tackling the Credit Assignment Problem in Reinforcement Learning-Induced Pedagogical Policies with Neural Networks

摘要

Intelligent Tutoring Systems (ITS) provide a powerful tool for students to learn in an adaptive, personalized, and goal-oriented manner. In recent years, Reinforcement Learning (RL) has shown to be capable of leveraging previous student data to induce effective pedagogical policies for future students. One of the most desirable goals of these policies is to maximize student learning gains while minimizing the training time. However, this metric is often not available until a student has completed the entire tutor. For this reason, the reinforcement signal of the effectiveness of the tutor is delayed. Assigning credit for each intermediate action based on a delayed reward is a challenging problem denoted the temporal Credit Assignment Problem (CAP). The CAP makes it difficult for most RL algorithms to assign credit to each action. In this work, we develop a general Neural Network-based algorithm that tackles the CAP by inferring immediate rewards from delayed rewards. We perform two empirical classroom studies, and the results show that this algorithm, in combination with a Deep RL agent, can improve student learning performance while reducing training time.