Source-linked AI summary
A GNN-Enhanced Reinforcement Learning Framework for Emergency Communications in ORAN-based Non-Terrestrial Networks
Md. Thouhidur Rahman, Mustapha Benjillali, Halim Yanikomeroglu, Samir Saoudi
TL;DR
Terrestrial stations can fail or struggle during disasters and other high-demand conditions, while ORAN-enabled NTNs offer flexible connectivity but add network-management complexity. This paper combines GNN topology modeling with Q-learning for adaptive emergency network control, reporting a successful action rate exceeding 80% while maintaining stable latency across learning configurations.
Problem
Terrestrial base stations often fail or struggle during disasters, rural conditions, and high data-rate demand, while distributed ORAN architectures complicate network management and resource optimization.
Source: §I · p. 1Method
The paper combines GNN topology modeling with Q-learning in an MDP to control ORAN functional splitting and resource management for emergency communications.
Source: §I · pp. 1–2Results
A successful action rate exceeding 80% was achieved while maintaining stable latency across varying learning configurations.
Source: §V · p. 6Takeaways & Limitations
The framework provides adaptive real-time control for ORAN-based NTN emergency networks.
Source: §V · p. 6
Abstract
from arXiv · showhide
During disaster scenarios and periods of extreme data demand in next-generation wireless communications, conventional terrestrial networks (TNs) often become unreliable or fail entirely, leading to critical service disruptions. In such contexts, non-terrestrial networks (NTNs) emerge as a promising solution to ubiquitous and resilient connectivity. Furthermore, the open radio access network (ORAN) paradigm facilitates network disaggregation and flexible functional splitting among its key components, namely the central unit (CU), distributed unit (DU), and radio unit (RU), which can be deployed across heterogeneous NTN platforms according to service requirements. However, this flexibility introduces significant challenges in terms of network complexity and real-time control. To address these challenges, this paper proposes an intelligent ORAN-enabled NTN framework for emergency communication scenarios. The proposed system leverages graph neural networks (GNNs) to model the dynamic network topology and employs a reinforcement learning (RL)-based Q-learning algorithm, formulated as a Markov decision process (MDP), to enable adaptive and real-time network control. In this framework, network nodes are treated as states, and optimal decisions are learned based on system dynamics. The spatial distribution of user equipment (UE) is modeled using an inhomogeneous Poisson point process (IPPP) with a rejection sampling technique, capturing realistic user density variations. Simulation results demonstrate that the proposed GNN-enhanced RL approach significantly improves network performance in terms of latency and service reliability, thereby enabling efficient and robust operation under emergency conditions.
I. INTRODUCTION
Terrestrial stations can fail or struggle during disasters and other high-demand conditions, while ORAN-enabled NTNs offer flexible connectivity but complicate network management. The paper proposes a GNN-enhanced RL framework for emergency network control and reports successful action rates of up to 90% while achieving target latency.
I. INTRODUCTION
- Terrestrial base stations often fail or struggle to provide reliable service during disasters, in rural areas, and under high data-rate demands, and.NTNs are presented as a promising source of flexible, resilient connectivity, while ORAN distributes CU, DU, and RU functions across heterogeneous platforms.Source: §I · p. 1
- The architecture deploys HAPS carrying open DU and RU functions first, then adds units according to user density and demand.A GNN digital twin monitors the physical components and supports control through GNN-enhanced RL algorithms.Source: §I · pp. 1–2
- UE distributions are modeled with an IPPP and rejection sampling based on damaged base-station locations to support emergency communications.The contribution also describes the resulting NTN architecture as efficient and scalable.Source: §I · p. 2
- The proposed framework combines a GNN network model with Q-learning in an MDP to optimize functional splits and resource management against target latency.The GNN represents NTN ORAN nodes and communication links to learn topology and link characteristics.Source: §I · p. 2
- Up to 90% successful action rates were achieved while the framework achieved the target latency.The introduction presents this as evidence of effective learning of optimal control policies and reliable, efficient network performance.Source: §I · p. 2
II. GNN FRAMEWORK FOR NTN ARCHITECTURE
The GNN framework represents the NTN as a graph of network entities and communication links, while modeling UE locations with a nonuniform distribution centered around the damaged base station.
II. GNN FRAMEWORK FOR NTN ARCHITECTURE
- The graph models NTN platforms, UEs, the CU gateway, core and data networks as nodes, and FSO and RF communications as edges.Node attributes include platform and ORAN-component types, position, neighboring nodes and link types.Source: §II · p. 3
- UE locations follow a two-dimensional Gaussian IPPP, with higher density near the damaged base station and decreasing density toward the area edges.HAPS is positioned above the damaged base station and a LEO relay above the HAPS; other platforms are placed according to UE spatial density.Source: §II · p. 3
- Rejection sampling accepts locations when funiform(xi, yi) < faccepted(xi, yi).The accepted PDF is the PDF used by the rejection-sampling method.Source: §II · p. 3
III. RL MODEL FOR NETWORK ADAPTATION
The RL model treats the network as a dynamic MDP and learns actions that adapt its topology and resource use. Its reward combines latency, UE-serving, and UAV-action criteria.
III. RL MODEL FOR NETWORK ADAPTATION
- The MDP agent observes the full graph and can add or remove nodes and links or update node positions under UE requirements and target-latency constraints.The network state evolves with user density and the agent’s actions.Source: §III · p. 3
- Latency above the target incurs a significant penalty, while latency below the target earns a reward.The latency reward uses average latency, target latency, and a penalty factor.Source: §III · p. 4
- Serving more UEs earns a positive reward, while reduced serving capability earns a negative reward.The UE-based reward compares served UEs with a baseline number of UEs.Source: §III · p. 4
- The total reward combines latency-based, UE-based, and action-based criteria.The component rewards assess target latency, served-UE counts, and DU/RU UAV changes.Source: §III · p. 4
- Removing DU or RU UAVs earns a positive reward when service efficiency improves; unnecessary additions or removals may be penalized.The action-based reward assigns a penalty when a DU and RU UAV are added and a reward when they are removed.Source: §III · p. 4
IV. GNN-ENHANCED RL TRAINING PERFORMANCE ANALYSIS
The training analyses show that higher learning rates improve latency compliance and successful-action rates, while user load and target latency shape reward trajectories.
IV. GNN-ENHANCED RL TRAINING PERFORMANCE ANALYSIS
- Q-learning observes the network state, selects node and link actions, updates positions and latencies, then evaluates actions using a reward.The state includes DU, RU, and UE counts; actions include adding or removing nodes and links.Source: §IV · p. 4
- The end-to-end graph includes 286 UEs, 31 UAVs, 27 O-RUs, 6 O-DUs, and one relay unit across NTN platforms.The paper contrasts this complete topology with the smaller network unit shown in Fig. 2.Source: §IV · p. 4, Figure 3 · p. 5, Figure 2 · p. 5
- Total training reward decreases and fluctuates more as UE numbers rise, while the 186-UE scenario converges in fewer than 100 episodes.With fewer users, redundant DU or RU UAVs can be removed, increasing action-based rewards.Source: §IV · pp. 4–5, Figure 4 · p. 5
- Lower target latencies yield higher total rewards, and rewards become more stable after approximately 400 episodes.The paper attributes the higher rewards to greater contrast between the average and target latencies.Source: Figure 5 · p. 5, §IV · p. 5
- As learning rates increase, latency exceeding the target decreases; average latency above target also declines as training episodes increase.The reported explanation is that higher learning rates help the agent learn the environment faster and reduce deviation from the target.Source: Figure 6 · p. 5, §IV · pp. 5–6
- Successful-action rates rise with learning rate, while actions are unsuccessful when average latency exceeds the target.The results attribute the increase to faster adaptation and more accurate decisions.Source: Figure 7 · p. 6, §IV · p. 6
V. CONCLUSION AND FUTURE WORK
The paper concludes that its ORAN-based NTN control system addresses emergency-network control challenges with GNN-enhanced RL. Its evaluation reports successful action rates above 80% and stable latency, while identifying several directions for future work.
V. CONCLUSION AND FUTURE WORK
- The proposed model achieves a successful action rate exceeding 80% while maintaining stable latency across learning configurations and observation conditions.The evaluation considers total reward, average latency, and successful action rate.Source: §V · p. 6
- Future work will optimize hyperparameters, improve generalization, and incorporate additional state representations.Source: §V · p. 6
- The authors plan baseline comparisons and integration of real-time UAV-assisted NTN data toward more efficient and scalable algorithms for practical deployment.Source: §V · p. 6