From Learning Analytics to Autonomous Agents: A Context-Aware Multimodal DRL Framework for Personalized Education

This paper introduces the Context-Aware Multi-Agent Deep Reinforcement Learning (CA-MA-DRL) framework, a transformative approach to personalized digital education that shifts the paradigm from passive learning analytics to autonomous, decision-making agents. The framework addresses the limitations of static, rule-based systems by integrating Multimodal Learning Analytics (MmLA) with advanced coordination mechanisms. By fusing heterogeneous data streams from Learning Management Systems (LMS), virtual classrooms, digital laboratories, and AR/VR platforms, the system constructs a rich Context-Aware Multimodal Representation that captures cognitive mastery, affective states, and behavioral engagement in real-time. The core architecture employs Student Agents and Teacher Agents that utilize Deep Q-Networks (DQN) and Actor- Critic architectures to learn optimal, personalized pedagogical policies. These agents engage in formalized negotiation protocols to balance learner agency with pedagogical integrity and institutional constraints. Incorporating a Human-inthe- loop oversight model through explainable AI (XAI), the framework ensures that instructors retain authority while benefiting from computational assistance. Comparative analysis across eight capability dimensions demonstrates that CAMA- DRL achieves 86% aggregate capability (69/80) compared to 29% for LLM Multi-Agent, 26% for Single-Agent DRL, and 18% for Rule-Based ITS, while resolving the accuracy-scalability trade-off through shared policy networks with meta-learning transfer. This digital-first foundation enables scalable and ethical personalization and provides a systematic pathway toward future hybrid cyber-physical and embodied educational AI systems.