Ilmu Komputer & AI editorial
Open AccessOA2026
Traffic-Adaptive Per-Hop Multipath Routing in Multi-Hop UAV Networks
A Dec-POMDP formulation and MAPPO-DM multi-agent reinforcement learning solution for on-time packet delivery in UAV-relayed MEC networks
Zhenyu Zhao; Tiankui Zhang; Xiaoxia Xu; Yuanpeng Zheng; Junjie Li; Wenjuan Xing· 2026· DOI 10.48550/arXiv.2608.25383
The core problem
In uncrewed aerial vehicle (UAV)-relayed mobile edge computing (MEC) networks, computation tasks generate traffic with diverse latency requirements and data sizes. Routing decisions therefore need to adapt to both traffic characteristics and changing network conditions. Compared with single-path routing, multipath routing is better suited to such heterogeneous traffic because it provides multiple forwarding options and enables flexible traffic splitting. However, conventional multipath routing usually splits traffic over predefined end-to-end paths, making it difficult to respond quickly to link fluctuations and topology changes in UAV networks. To address this issue, the authors propose a traffic-adaptive per-hop multipath routing method for multi-hop UAV networks, in which each UAV dynamically distributes traffic among multiple candidate next hops. The routing problem is formulated to improve the on-time packet delivery ratio while reducing the packet loss ratio, and is modeled as a decentralized partially observable Markov decision process (Dec-POMDP).
Innovation
Simulation results show that MAPPO-DM outperforms the baseline methods and maintains robust performance under various network conditions. The evaluation focuses on two primary metrics: on-time packet delivery ratio and packet loss ratio. Across the tested scenarios, MAPPO-DM achieves higher on-time delivery and lower packet loss compared with the baselines. The robustness of the approach is demonstrated by consistent performance across different network conditions, indicating that the per-hop, traffic-adaptive splitting strategy generalizes beyond a single operating point. The abstract does not report specific numerical values for these improvements; the source only states that MAPPO-DM outperforms baselines and remains robust under varied conditions.
In uncrewed aerial vehicle (UAV)-relayed mobile edge computing (MEC) networks, computation tasks generate traffic with diverse latency requirements and data sizes. Routing decisions therefore need to adapt to both traffic characteristics and changing network conditions. Compared with single-path routing, multipath routing is better suited to such heterogeneous traffic because it provides multiple forwarding options and enables flexible traffic splitting. However, conventional multipath routing usually splits traffic over predefined end-to-end paths, making it difficult to respond quickly to link fluctuations and topology changes in UAV networks. To address this issue, the authors propose a traffic-adaptive per-hop multipath routing method for multi-hop UAV networks, in which each UAV dynamically distributes traffic among multiple candidate next hops. The routing problem is formulated to improve the on-time packet delivery ratio while reducing the packet loss ratio, and is modeled as a decentralized partially observable Markov decision process (Dec-POMDP).
The routing problem is cast as a Dec-POMDP, which captures the decentralized nature of per-hop forwarding decisions under partial observability of network state. To solve it, the authors develop a multi-agent reinforcement learning (MARL) algorithm termed Multi-Agent Proximal Policy Optimization with Dirichlet Modeling (MAPPO-DM). MAPPO-DM follows the centralized-training-and-decentralized-execution (CTDE) framework: during training, a centralized critic can access global information, while during execution each UAV acts only on its local observations. Continuous traffic-splitting actions are modeled using a Dirichlet distribution, which naturally represents a probability simplex over candidate next hops. This allows each UAV to allocate fractions of traffic across multiple next hops in a single decision step, rather than selecting a single path.
Why it matters
The key insight of this work is that shifting from end-to-end path selection to per-hop traffic splitting enables faster adaptation to link fluctuations and topology changes in multi-hop UAV networks. By modeling continuous splitting actions with a Dirichlet distribution, MAPPO-DM avoids the combinatorial action space that would arise from enumerating discrete path combinations, while still allowing flexible allocation across multiple candidate next hops. The Dec-POMDP formulation aligns with the practical constraint that each UAV has only partial observability of the network, and the CTDE framework permits stable training without requiring centralized execution. The reported robustness across network conditions suggests that the learned policy does not overfit to a specific topology or traffic pattern. However, the source does not provide detailed ablation studies, convergence analysis, or comparisons against a wide range of MARL baselines in the abstract; these details would be needed to fully assess the method's advantages. Overall, the approach offers a promising direction for traffic-adaptive routing in UAV-relayed MEC networks where heterogeneous latency requirements and dynamic topology are the norm.
Who should read this
CS practitioners and researchers
Opening member content…