Abstract
Low-Altitude Economy Networks (LAENets) are emerging as a critical infrastructure for supporting low-altitude services through integrated air-ground systems. By leveraging Unmanned Aerial Vehicles (UAVs) equipped with Mobile Edge Computing (MEC) capabilities, these networks can offload intensive computing tasks from ground devices to aerial nodes. However, achieving low-carbon operations in multi-UAV-assisted MEC networks presents significant challenges, specifically in multi-dimensional modeling and multi-objective coupled optimization. This paper introduces a two-stage framework to address these hurdles. First, we propose HybridRAG, a Retrieval-Augmented Generation (RAG) framework that combines KeywordRAG, VectorRAG, and GraphRAG to empower Large Language Model (LLM) agents in formulating precise optimization problems. Second, we introduce the Double Regularization Diffusion-enhanced Soft Actor-Critic ($R^2DSAC$) algorithm to solve these problems efficiently. Our approach incorporates diffusion entropy and action entropy regularization, alongside dynamic neuron masking, to optimize performance while minimizing the carbon footprint of the AI training process itself.
Introduction
The "Low-Altitude Economy" represents a burgeoning sector where fleets of UAVs provide services ranging from logistics to environmental surveillance. To meet the low-latency and high-computation demands of these services, UAVs are increasingly integrated into Mobile Edge Computing (MEC) architectures. While this integration offers flexibility and resilience, the environmental impact of these networks is a growing concern. Developing "green" LAENets requires balancing energy consumption, signal interference, and flight physics.
Two primary gaps hinder the implementation of low-carbon LAENets:
- The Modeling Gap: Formulating the mathematical models for multi-UAV systems is complex. Human engineers must account for intricate relationships between flight dynamics, communication constraints, and energy models, a process prone to error.
- The Optimization Gap: The resulting multi-objective optimization problems are typically NP-hard. Traditional reinforcement learning (RL) methods often struggle with high-dimensional action spaces, leading to suboptimal resource allocation and flight trajectories.
Methodology
1. HybridRAG for Automated Problem Formulation
To bridge the modeling gap, we developed a HybridRAG-based LLM agent. Unlike standard RAG systems that rely solely on semantic similarity, HybridRAG integrates three distinct retrieval mechanisms to ensure technical accuracy:
- VectorRAG: Provides general semantic context through traditional embedding-based search.
- KeywordRAG: Ensures precise matching for domain-specific technical terminology.
- GraphRAG: Utilizes a knowledge graph (implemented via Neo4j) to capture structural relationships between variables. For instance, it explicitly links "UAV velocity" to "propulsion energy consumption," ensuring the LLM understands the physical dependencies required for accurate mathematical modeling.
This hybrid approach allows the LLM to act as an "auxiliary engineer," translating high-level requirements into rigorous optimization formulations.
2. $R^2DSAC$ for Low-Carbon Optimization
Once the optimization problem is formulated, we employ the $R^2DSAC$ algorithm to find the optimal policy for UAV trajectory and resource management. This algorithm introduces several innovations:
- Diffusion Policy: Rather than using a standard neural network for the actor, we implement a diffusion model. This allows the agent to represent complex, multi-modal action distributions, which is essential for navigating the vast search space of multi-UAV networks.
- Double Regularization: To prevent the model from converging on suboptimal "local minima," we apply both diffusion entropy and action entropy regularization. This encourages exploration and prevents the policy from becoming overconfident in inefficient strategies.
- Dynamic Pruning (Green AI): Recognizing that AI training itself consumes significant energy, we implement a dynamic masking module. By identifying and masking unimportant neurons in the actor network during training, we reduce the computational overhead and associated carbon emissions without sacrificing performance.
Experimental Evaluation
Dataset and Environment
Our framework was evaluated using an expert knowledge base curated from IEEE Xplore, focusing on MEC, carbon reduction, and task offloading. The simulation environment modeled a $1000m \times 1000m$ area with 2 UAVs and 10 ground users, testing the agent's ability to manage trajectories and offloading decisions.
Key Results
- Retrieval Performance: HybridRAG demonstrated superior accuracy in technical context retrieval. It achieved a Context Utilization (CU) score of 80.2%, significantly outperforming the 75.8% achieved by standard Vector+Keyword RAG configurations.
- Optimization Efficiency: The $R^2DSAC$ algorithm achieved a 64% improvement in total rewards compared to the baseline Soft Actor-Critic (SAC) algorithm. It also exhibited faster convergence rates than Proximal Policy Optimization (PPO).
- Sustainability: Through dynamic pruning, the framework maintained high-performance flight trajectories while limiting training-related carbon emissions to approximately 70.3 grams, demonstrating the feasibility of "Green AI" for environmental optimization.
Conclusion and Future Work
This research demonstrates that combining structural retrieval (HybridRAG) with advanced diffusion-based reinforcement learning ($R^2DSAC$) can effectively address the complexities of low-carbon LAENets. By positioning LLMs as modeling assistants and optimizing the carbon cost of the AI itself, we move closer to sustainable autonomous aerial networks.
Future research will focus on refining prompt engineering through automated optimization, extending the framework to decentralized multi-agent diffusion models for larger UAV swarms, and validating the system on physical hardware under real-world atmospheric conditions.
Summary updated