A LATENCY-AWARE AND LOAD-BALANCED VIRTUAL NETWORK EMBEDDING FRAMEWORK USING GRAPH ATTENTION NETWORKS AND HYBRID PROXIMAL POLICY OPTIMIZATION

Abstract

Virtual Network Embedding (VNE) is crucial for efficiently mapping virtual requests onto physical infrastructures in large-scale networks such as 5G, IoT, and edge-cloud systems. However, existing approaches often struggle to adapt to dynamic network states, minimize latency, balance resource loads, and achieve rapid and stable learning. To address these challenges, this study introduces ALIVE (Adaptive Latency-aware Intelligent Virtual Network Embedding), an enhanced VNE framework incorporating four key innovations. First, latency sensitivity and load balancing are integrated as core optimization objectives to enable low-delay path selection while preventing resource bottlenecks. Second, Graph Attention Networks (GATs) are utilized to improve node representation by dynamically weighting neighbor importance, which is particularly beneficial in heterogeneous and time-varying network topologies. Third, Hybrid Proximal Policy Optimization (HPPO) is employed to stabilize the learning process, accelerate convergence, and enhance robustness. Finally, Meta-Gradient Learning is incorporated for adaptive learning rate control, allowing autonomous tuning of learning dynamics based on reward feedback. Simulation results demonstrate that ALIVE significantly outperforms existing state-of-the-art methods by achieving a higher acceptance ratio, lower latency, improved load balance, stable convergence, and superior adaptability under real-time network conditions.