Petri Nets and Hierarchical Reinforcement Learning for Personalized Student Assistance in Serious Games

Abstract

Adaptive serious games offer a new frontier for education, especially in complex topics. However, optimal methods for in-game adaptation are still being explored to address challenges such as limited educator resources, unpredictable or limited data, or complicated implementation procedures. This work offers an adaptable framework for personalized student assistance and directing within an adaptive serious game using reinforcement learning and Petri nets. Our proposed framework can be built upon by serious game developers and researchers to create adaptive serious games for improving student learning in other domains. Building on prior work, we address the challenge of adaptive in-game content through Petri net player modelling and a multi-agent deep reinforcement learning approach to gradually learn optimal personalized assistance. Finally, we provide proof-of-concept training performance for our proposed agent using a student simulation, demonstrating that the proposed hierarchical reinforcement learning approach offers significantly (effect size r = 0.8101) improved training performance over a tabular, single-agent approach.