Dynamic Service Function Chain (SFC) Deployment for Autonomous Intelligent Systems

Abstract

Autonomous Intelligent Systems (AIS), such as autonomous vehicles, smart drones, and robotic swarms, are becoming increasingly prevalent in next-generation intelligent infrastructures. However, AIS environments are characterized by highly dynamic network conditions, severe device heterogeneity, strict end-to-end latency requirements, and strong privacy constraints due to distributed multimodal sensing data. These challenges render traditional Service Function Chain (SFC) deployment approaches, especially centralized or slow-adapting methods, unsuitable for real-world AIS scenarios. To address these challenges, we propose MetaFedDRL, a meta-federated reinforcement learning framework designed for fast, privacy-preserving, and communication-efficient SFC deployment in large-scale AIS networks. MetaFedDRL integrates Model-Agnostic Meta-Learning (MAML), Proximal Policy Optimization (PPO), and Asynchronous Federated Learning (AFL) to learn a generalizable policy initialization across diverse environments. This enables rapid adaptation to new conditions with only a few local gradient updates, significantly reducing convergence time. The framework exchanges only lightweight gradient information, thereby preserving data privacy and minimizing communication overhead. Extensive simulations under dynamic and heterogeneous AIS settings show that MetaFedDRL achieves higher SFC acceptance rates, lower end-to-end latency, reduced average resource consumption, and improved deployment fairness compared to state-of-the-art baselines.