Intelligent Virtual Agents in Primary Education: Investigating Adaptive Mechanisms in a Digital Reading Intervention Using Reinforcement Learning

The utilization of technology in education offers numerous advantages for accelerating and enhancing a process that is fundamental throughout one's entire life: learning. In the pursuit of supporting this critical process, technology has been employed in various ways to assist struggling learners. This cumulative dissertation presents interdisciplinary research at the intersection of computer science and psychology, investigating the development and evaluation of an Intelligent Virtual Agent (IVA) embedded in an educational mobile application for primary education, focusing on the integration of adaptive mechanisms through Artificial Intelligence (AI) in a step-by-step process. Thus, this work has two main contributions. First, it investigates the human-centered design and impact of an iteratively conceptualized mobile reading application for primary students, implementing an IVA. This digital reading training has been developed within an interdisciplinary research project setting. It adheres to established usability principles and is based on a validated analogue reading intervention, employing theoretical knowledge of psychological concepts to integrate a feedback system. The digital reading training has been empirically validated with primary students in a study with quasi-experimental pretest/posttest design, offering evidence that the digital intervention benefits word reading of low-skilled readers. Further, usability and feasibility evaluations confirmed its suitability for primary school learners. Second, the educational mobile application and the IVA's behavior are further extended by integrating adaptivity through AI techniques, namely Reinforcement Learning (RL). Though RL has proven to be a suitable method for integrating adaptivity in educational contexts, its application in primary education is rather limited. This work aims to address this research gap by examining the use of such adaptive mechanisms to support students in the learning process. Employing RL methods allows the IVA to adapt its feedback behavior to each child's learning needs. Through several real-world field studies and a long-term evaluation of the RL-powered IVA in a classroom setting, this work indicates that the integration of RL-based adaptivity in this educational context is technically feasible, however, empirical results revealed no significant advantage over the mobile app version with the non-adaptive IVA. These findings underscore that while RL offers a promising framework for personalization, its educational effectiveness heavily depends on contextual factors such as domain complexity, action space, and baseline quality. While the digital reading training has generally proven effective for low-proficient readers, the results of this work highlight the importance of grounding adaptive technologies in both theory and real-world research. In doing so, this dissertation advances the field of educational mobile applications, AI-powered IVAs, and RL for education by offering valuable insights and practical implications for future research on designing and evaluating educational technologies in authentic, real-world contexts to fully understand and leverage the potential of AI-based IVAs in supporting learners' needs.