A context-aware multimodal multi-agent deep reinforcement learning framework for autonomous personalized education

Abstract

This paper introduces the Context-Aware Multi-Agent Deep Reinforcement Learning (CA-MA-DRL) framework for personalised digital education, shifting from passive analytics to autonomous decision-making agents. The framework integrates Multimodal Learning Analytics with advanced coordination mechanisms, fusing heterogeneous data from LMS platforms, virtual classrooms, and AR/VR environments to construct context-aware representations capturing cognitive, affective, and behavioural states. Student and Teacher Agents employ DQN and Actor-Critic architectures with formalised negotiation protocols, while Human-in-the-loop oversight ensures instructor authority through explainable AI. This is a conceptual architecture paper: we contribute a fully specified design, a reference implementation configuration grounded in the authors’ previously validated CA-MA-DRL deployment, and a pre-specified multi-phase evaluation protocol, rather than a completed empirical study. A structured eight-dimension capability assessment—an analytical design comparison rather than a measurement of performance—indicates a substantially higher aggregate capability for CA-MA-DRL than for the LLM Multi-Agent and Rule-Based ITS baselines, with full per-dimension scores reported in the paper. The framework is designed to address the accuracy-scalability trade-off through shared policy networks with meta-learning transfer.