Proximal Policy Optimization for AI-Orchestrated SD-WAN in Underserved Regions

This paper explores the scenario of rural connectivity by leveraging Software-Defined Wide Area Network (SDWAN) technology. We introduce an AI-driven Fabric SD-WAN architecture for rural connectivity, in which distributed edge agents and a central orchestrator cooperatively manage multi-access links (LEO satellite and LTE). A core challenge is dynamic tunnel selection in rural and remote environments. While prior works used deterministic heuristics and Deep Q-Networks (DQN), policy-based methods remain unexplored in this domain. We propose a tailored Proximal Policy Optimization (PPO) module within the fabric control plane that ingests real-time telemetry (tunnel status, bandwidth, traffic class) and outputs routing policies. In a realistic rural topology, our PPO-fabric achieves up to 98 % Connection Reliability Index, outperforming random, deterministic, and DQN baselines.