M2-MADRL: A multi-mode routing optimization strategy based on MADRL for Edge Ad Hoc Networks

This paper addresses the challenges of dynamic routing in Edge Ad Hoc Networks (EANETs), such as high mobility, complex topologies, and diverse service requirements, by proposing M2-MADRL, a Multi-Mode routing framework based on Multi-Agent Deep Reinforcement Learning. The study focuses on three key aspects: first, designing a multi-mode routing architecture that enables nodes to dynamically switch between routing protocols (e.g., AODV, DSDV) and adaptive parameter adjustments, adapting to varying network scales and mobility levels. Second, integrating an enhanced MADRL model based on the Centralized Training with Decentralized Execution (CTDE) framework, which uses local value networks for individual node optimization and a global value network to improve overall network performance, resolving conflicts between centralized control and EANETs’ distributed nature. Third, introducing a double-layer Semantic Self-Attention Mechanism (SSAM) to help agents accurately interpret heterogeneous network states and context-aware actions, accelerating convergence and enhancing decision-making precision. Experiments show M2-MADRL outperforms MADDPG, DE-MADDPG, and RTHop in Packet Delivery Rate (PDR), Effective Throughput (ET), and Average Transmission Delay (ATD), demonstrating superior adaptability in EANETs. For example, in 150-node networks, it averagely demonstrates 11.9% higher PDR, 23.0% higher ET, and 44.6% lower ATD than the classic MADDPG method.