TL;DR
This paper introduces a novel deep learning-based routing approach for aeronautical ad-hoc networks (AANETs) that leverages real flight data. It tackles the challenge of highly dynamic network topologies by enabling aircraft to make routing decisions using only local information, achieving performance close to optimal global routing and approaching Pareto fronts for multi-objective optimization.
What Problem Does This Paper Address?
Aeronautical ad-hoc networks (AANETs) are envisioned to provide global internet connectivity by using commercial airplanes as relays. However, the high speed of aircraft leads to rapidly changing network topologies, making traditional routing protocols inefficient. Existing position-based routing methods struggle with "communication voids" (where no neighbor is closer to the destination), and even advanced methods often require global network knowledge or struggle with non-convex multi-objective optimization problems. This paper aims to develop a routing solution that is efficient, adaptable to dynamic environments, and can optimize multiple performance metrics (delay, capacity, lifetime) using only local information.
Methods
The core of the proposed solution is a Deep Neural Network (DNN) trained on historical flight data.
- Single-Objective Routing (Minimum Delay): A DNN is trained to map local geographic information (positions of the current node, neighbors, and destination) to the minimum end-to-end (E2E) delay from neighboring nodes to the destination. This allows a forwarding node to determine the optimal next hop without needing global topology information. The training leverages the predictable mobility patterns of commercial flights.
- Multi-Objective Routing (Delay, Capacity, Lifetime): The approach is extended to simultaneously optimize delay, path capacity, and path lifetime. This involves a more complex DNN that predicts these three metrics for neighboring nodes. The problem is framed using an epsilon-constraint method, and the DNN is trained to estimate the required values for these constraints.
- Feedback Mechanism: For multi-objective routing, a feedback mechanism is introduced where next-hop candidates also provide their estimated performance metrics, allowing the forwarding node to make more informed decisions.
- Offline Training, Online Decision: The DNNs are trained offline using historical data. Once trained, the DNN is deployed on each aircraft, enabling real-time, distributed routing decisions based solely on local observations.
Datasets
- Real Flight Data: The paper utilizes real flight data collected over the North Atlantic Ocean and the European continent. This data includes timestamps, latitude, longitude, altitude, speed, and heading for flights.
- Synthetic Flight Data: To augment the real data and create diverse training and testing environments, synthetic flight data is generated by randomly shifting the real flight data along the timeline. This helps reflect mismatches between training and testing environments.
Key Insights and Ideas
- Exploiting Predictable Mobility: The regular and predictable flight paths of commercial aircraft are a key insight. This regularity allows for effective offline training of DNNs, enabling them to learn routing strategies that generalize well.
- Local Information Sufficiency: The research demonstrates that by intelligently using DNNs trained on historical data, routing decisions can be made effectively using only local geographic information, overcoming the limitations of protocols requiring global state.
- DNNs as "Topology Forecasters": The DNN acts as a surrogate for global topology knowledge, inferring necessary information about future network states and path performance from local observations.
- Near-Pareto Optimality: The multi-objective approach, even when relying on local information, can achieve performance close to the true Pareto front obtained with global knowledge, offering a practical trade-off between optimality and feasibility.
Key Results
- Single-Objective Performance: The proposed DL-aided routing significantly outperforms existing position-based routing protocols (like GPSR and GLSR) in terms of E2E delay, path capacity, and path lifetime. It achieves performance close to optimal routing that relies on global link information.
- Multi-Objective Performance: The DL-aided multi-objective routing algorithm can find paths that approach the Pareto front, demonstrating a good trade-off between delay, capacity, and lifetime. It also guarantees the specified lifetime constraints, which other benchmark algorithms sometimes fail to do.
- Computational Efficiency: While the multi-objective DNN has higher complexity, the one-shot evaluation time for both single- and multi-objective routing is very low (less than 1 ms), making it suitable for real-time decision-making.
Limitations and Future Work
- Flight Density: The paper acknowledges that AANETs might struggle in regions with low flight density.
- MOO Complexity: Multi-objective optimization introduces higher computational complexity and resource demands.
- Future Work:
- Integrating satellites into AANETs for truly global coverage.
- Further reducing the complexity of multi-objective optimization.
- Investigating how to handle situations with less predictable flight patterns or lower flight densities.
Summary updated