Introduction
Urban Air Mobility (UAM), characterized by the deployment of "flying cars," necessitates high-fidelity environmental perception to ensure operational safety. However, individual aerial vehicles often face constraints in onboard sensing and computational capacity. While Federated Learning (FL) offers a pathway for collaborative intelligence without raw data sharing, traditional centralized FL architectures introduce significant communication bottlenecks and single points of failure in dynamic, high-speed aerial environments.
Fully decentralized learning, where vehicles interact only with local neighbors, emerges as a robust alternative. Nevertheless, this approach is susceptible to "consensus" failures: if the network topology is suboptimal or communication delays are excessive, local AI models may fail to converge. This paper proposes a novel framework that leverages Reconfigurable Intelligent Surfaces (RIS) to actively reshape the network's physical communication topology, ensuring reliable decentralized learning.
Methodology: RIS as a Network Architect
The core innovation lies in treating the wireless environment not as a fixed constraint, but as a controllable variable. By adjusting the phase shifts of RIS elements, the system can "construct" new long-range links or "deconstruct" redundant ones to optimize the network structure for learning.
Topological Criteria
The authors utilize spectral graph theory to define the "health" of the learning network based on the eigenvalues of its Laplacian matrix:
- Minimizing $\lambda_{max}$: Reducing the largest eigenvalue enhances the network's tolerance to transmission delays, preventing divergence when the connection lags.
- Maximizing $\lambda_2$: Increasing the second smallest eigenvalue (algebraic connectivity) accelerates the rate at which distributed models reach consensus.
Optimization via DDPG
Determining optimal RIS phase shifts in real-time is computationally intensive. To address this, the researchers developed a Deep Deterministic Policy Gradient (DDPG) reinforcement learning agent. This agent learns to dynamically adjust the RIS to maximize a reward function tied to transmission rates and delay-tolerance thresholds.
Key Insights and Results
A significant theoretical finding is the "singleton neighbor structure" advantage, where sparse, non-overlapping local connections can paradoxically improve global network conductance. This shifts the view of RIS from a Physical Layer (Layer 1) signal booster to a Network Layer (Layer 2) topology controller.
Experimental evaluations using a Multi-view Vehicle Dataset (5,000 samples) across various architectures (MobileNet, DenseNet) demonstrated substantial gains:
- Enhanced Stability: The RIS-revised topology maintained convergence at transmission delays up to 0.28s, whereas the initial topology failed at 0.27s.
- Superior Accuracy: Distributed learning on the optimized topology consistently outperformed standard ring or random configurations.
- Robustness: While centralized FL achieved higher peak accuracy, it required significantly lower delays ($\tau_{max} = 0.19s$), making the decentralized RIS approach more viable for real-world aerial scenarios.
Conclusion
By integrating RIS into the topology control loop, this research provides a robust framework for decentralized learning in UAM. While challenges remain regarding fast topological changes and the overhead of channel state information (CSI) estimation, this work marks a pivotal step toward self-organizing, intelligent aerial networks.
Summary updated