Adaptive Knowledge Tracing Through Multi-Agent Reinforcement Learning: A Framework for Personalized Learning Path Optimization

Abstract

Personalized learning systems require accurate modeling of student knowledge states and adaptive curriculum sequencing to optimize learning outcomes. Traditional knowledge tracing approaches such as Bayesian Knowledge Tracing and recent deep learning methods like Deep Knowledge Tracing face limitations in coordinating multiple pedagogical objectives simultaneously. This paper proposes a novel multi-agent reinforcement learning framework that decomposes the adaptive learning problem into specialized agents responsible for knowledge estimation, content selection, difficulty calibration, and engagement optimization. The framework employs differentiable inter-agent communication protocols inspired by DIAL architecture and dueling network structures for robust Q-value estimation. Experimental results on the ASSISTments dataset demonstrate that the proposed MARL framework achieves 23.7% improvement in prediction accuracy (AUC 0.847 vs. 0.685) and 31.2% reduction in learning time compared to Deep Knowledge Tracing baselines, while maintaining high student engagement levels. The multi-agent coordination mechanism enables effective decomposition of complex educational objectives and provides interpretable insights into personalized learning path optimization strategies.