Hierarchical Online Optimization Approach for IRS-enabled Low-altitude MEC in Vehicular Networks
Abstract
This paper addresses the challenge of computation offloading in urban vehicular networks where line-of-sight (LoS) links are frequently obstructed. We propose a Hierarchical Online Optimization Approach (HOOA) that leverages UAV-mounted edge servers and Intelligent Reflecting Surfaces (IRS) to enhance connectivity. The problem is formulated as a Stackelberg game, decoupling discrete offloading decisions from continuous trajectory and resource control. We introduce GDMTD3, a novel algorithm combining Generative Diffusion Models with Twin Delayed Deep Deterministic Policy Gradient (TD3), to optimize high-dimensional action spaces. Experimental results show that HOOA significantly reduces task latency and energy consumption compared to conventional deep reinforcement learning methods.
Introduction
Modern vehicular applications, such as autonomous driving and cooperative perception, demand high computational power. While Low-Altitude Wireless Networks (LAWNs) utilizing UAVs as mobile edge computing (MEC) servers provide a flexible solution, urban "canyons" often block direct communication links between vehicles and aerial servers. Intelligent Reflecting Surfaces (IRS) offer a promising remedy by creating virtual LoS paths through passive beamforming. However, the joint optimization of task offloading, UAV trajectory, IRS phase shifts, and resource allocation constitutes a complex mixed-integer non-linear programming (MINLP) problem that is NP-hard and highly coupled across time.
System Architecture
The proposed system consists of vehicles, a base station (BS) with an MEC server, a UAV-mounted MEC server, and IRS units deployed on buildings and the UAV itself.
As shown in Figure 1, the network must manage multiple link types, including Vehicle-to-UAV (V2U), Vehicle-to-BS (V2BS), and reflected links via IRS. The goal is to minimize a weighted sum of task completion delay and energy consumption by dynamically adjusting the system's configuration in response to vehicle mobility.
The HOOA Framework
To solve the MINLP problem, we employ a hierarchical approach modeled as a Stackelberg game. This structure allows the system to handle discrete and continuous variables separately, improving computational efficiency.
As illustrated in Figure 2, the framework is divided into two distinct levels:
- Leader Level (Continuous Decisions): The MEC servers (UAV and BS) act as leaders. We propose the GDMTD3 algorithm, which integrates a Generative Diffusion Model (GDM) into the TD3 framework. The GDM serves as a powerful action refiner, enabling the agent to explore the high-dimensional space of UAV trajectories and IRS phase shifts more effectively than standard Gaussian noise-based exploration.
- Follower Level (Discrete Decisions): Vehicles act as followers, deciding where to offload their tasks. This is handled via a many-to-one matching mechanism, where vehicles and servers maintain preference lists based on local utility (latency vs. energy), ensuring a stable and feasible association.
- Resource Allocation: To further reduce the learning burden on the AI agent, computation resource allocation is solved analytically using Karush-Kuhn-Tucker (KKT) conditions, providing a closed-form solution once the trajectory and offloading decisions are fixed.
Experimental Evaluation
The HOOA framework was evaluated in a simulated 1000m × 1000m urban environment with 10 vehicles following a Gauss-Markov mobility model. Key findings include:
- Performance Gains: HOOA achieved a 2.5% reduction in average task completion delay and a 3.1% reduction in energy consumption compared to state-of-the-art DRL baselines like SAC and standard TD3.
- Convergence and Stability: The inclusion of the Diffusion Model led to smoother convergence and higher cumulative rewards, demonstrating superior stability in complex environments where standard DRL often fluctuates.
- Scalability: The approach remained robust under varying vehicle densities and task loads, proving its potential for real-world urban deployment.
Conclusion
The HOOA framework effectively addresses the challenges of urban MEC by combining hierarchical game theory, generative AI, and classical optimization. By utilizing IRS to bypass blockages and Diffusion Models to refine complex trajectories, the system achieves significant efficiency gains. Future work will focus on multi-UAV coordination and optimizing the computational overhead of the diffusion models for real-time deployment on resource-constrained aerial platforms.
Summary updated