Spatiotemporal Attention-Based Deep Reinforcement Learning for Joint Routing and Resource Management in Wireless Ad-Hoc Networks

Establishing high-quality communication routes in wireless ad-hoc networks remains challenging due to the strong coupling between routing strategies and resource management. To enable energy-efficient and robust networking, this paper proposes a novel cross-layer optimization framework based on spatiotemporal attention mechanism for joint routing and power management. The framework employs a deep reinforcement learning (DRL) method with hop-by-hop optimization to tackle the inherent partially observable Markov Decision Problem (POMDP) encountered during route establishment. At each node, a graph attention network (GAT) captures the irregular topology and dynamic interference relationships within its local neighborhood, while a Transformer encoder module models temporal evolution of channel and traffic conditions from historical observations. Priority experience replay is incorporated to improve sample efficiency and accelerate convergence. By effectively integrating spatial graph information and temporal context, the method significantly extends effective observability, thereby enabling superior interference mitigation and more efficient power management. Numerical simulations demonstrate that the proposed spatiotemporal attention-based approach outperforms conventional and DRL-based baselines in bottleneck throughput, transmission delay, and power efficiency. These results confirm the framework’s effectiveness, robustness and strong suitability for green communication in highly dynamic and resource-constrained wireless networks.