Designing Low-Resource Pedagogical Intervention Controllers: A Comparative Study of Compact Neural and Neuroevolutionary Architectures

Abstract

In the context of Massive Open Online Courses (MOOCs), data-driven models play a vital role in assisting students through personalized interventions. Earlier works have focused on the prediction of students’ performance and dropouts. However, the design of efficient decision-making controllers has received less attention. In this paper, the selection of pedagogical intervention is formulated as a low-resource controller design. A comparative study of traditional machine learning models, compact multilayer perceptron, and sparse neural controllers ob tained through neuroevolutionary search is proposed. In addition, a novel framework is proposed as a combination of NEAT and weight-agnostic neural networks. In the experiments, a large scale educational dataset is used. It is observed that compact neural models achieve high performance with reduced model sizes. Moreover, sparse neuroevolutionary controllers achieve extreme compression with acceptable performance. Thus, the need for balancing efficiency in the design of scalable intelligent educational systems is highlighted.

Keywords