Hierarchical Model Selection and Control for Latency-Energy Optimization in MEC-Assisted Vehicular Networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Engineering

Background:

• Multi-access edge computing (MEC) offloads tasks in vehicular networks.
• Current methods often use fixed AI models, ignoring selection's impact on performance.
• Optimizing AI model selection is crucial for MEC-assisted vehicular networks.

Purpose of the Study:

• To propose a hierarchical model selection and control (HMSC) framework for MEC-assisted vehicular networks.
• To jointly optimize latency, energy consumption, and service reliability through adaptive AI model selection.
• To enhance computational offloading efficiency in dynamic vehicular environments.

Main Methods:

• Developed a hierarchical framework using deep reinforcement learning (DRL).
• Integrated a vehicle-layer component (MAPPO) for communication and energy accounting.
• Employed a centralized MEC-layer agent (SAC) for adaptive AI model selection and resource allocation under SDN.
• Defined a composite objective integrating latency, energy consumption, and deadline violation penalty.

Main Results:

• HMSC significantly reduced end-to-end (E2E) latency compared to static and non-hierarchical DRL baselines.
• Achieved a higher deadline satisfaction ratio (DSR) under constrained uplink throughput and varying traffic loads.
• Demonstrated load-aware policy favoring high-fidelity models under light load and lightweight models under congestion.
• Quantified inference-quality implications of adaptive selection using YOLOv5 accuracy.

Conclusions:

• Coordinated MEC-side control of AI model selection and computation offers a robust latency-energy trade-off.
• The HMSC framework effectively manages computational resources for MEC-assisted vehicular networks.
• Adaptive AI model selection is key to enhancing performance and reliability in vehicular edge computing.