Introduction
Mobile Edge Computing (MEC) has emerged as a critical technology for supporting resource-constrained Internet-of-Things devices (IDs) by bringing computational power to the network edge. Mounting MEC servers on Unmanned Aerial Vehicles (UAVs) adds a layer of flexibility, allowing for dynamic coverage and improved Line-of-Sight (LoS) communication. However, optimizing these networks is inherently complex due to the tight coupling between the UAV's trajectory and the offloading decisions of ground devices.
Existing solutions often rely on centralized control, which fails to scale, or rigid multi-agent systems that lack the coordination necessary to handle fluctuating device populations. This paper introduces Cooperative Multi-Agent Deep Deterministic Policy Gradient (C-MADDPG), a framework designed to enable autonomous coordination and seamless scalability in UAV-aided MEC environments.
System Model and Problem Formulation
In the considered scenario, a UAV serves a dynamic set of ground IDs. The network's efficiency depends on three interdependent variables: the UAV’s 3D trajectory, the allocation of its limited CPU resources, and the local offloading decisions of each ground device. The primary objective is to minimize the total energy consumption of the network while meeting the computational demands of all users.
The C-MADDPG Framework
The core innovation of C-MADDPG lies in its ability to let agents learn their own communication protocols rather than relying on human-designed messages. This is achieved through a Dual-Actor Architecture:
- Message Actors: These generate "message actions"—encoded vectors that encapsulate an agent's local state and requirements.
- Solution Actors: These take the received messages as side information to decide on physical actions, such as the UAV's next movement vector or an ID's offloading ratio.
To manage a varying number of IDs, the UAV agent utilizes a Graph Attention Network (GAT). By treating the network as a graph where the UAV is the central node, the GAT allows the UAV to "attend" to the most relevant devices regardless of the total population. This decoupling of the neural network architecture from the number of users solves the scalability issues common in fixed-input models.
Key Results and Insights
The framework was evaluated using a custom simulation environment incorporating Gauss-Markov mobility models and realistic air-to-ground channel characteristics.
- Superior Energy Efficiency: C-MADDPG significantly outperformed vanilla MADDPG and non-cooperative baselines, achieving energy savings comparable to ideal centralized systems but with significantly lower overhead.
- Robust Scalability: A model trained on only 5–10 IDs demonstrated negligible performance loss when deployed in a network with 30 IDs. This generalization is attributed to the GAT-based message aggregator and a random masking strategy used during training.
- Autonomous Protocol Emergence: The system successfully learned to encapsulate useful statistics into messages without manual engineering, allowing for faster adaptation when the number of IDs changed mid-mission.
Conclusion
C-MADDPG represents a significant step toward autonomous, scalable management of UAV-aided MEC networks. By combining learned communication with attention-based aggregation, the framework balances local autonomy with global coordination. Future work will focus on extending this model to multi-UAV swarms and addressing potential communication bottlenecks in spectrum-congested environments.
Summary updated