Technology Category
Search and Optimization: Learning to SearchMachine Learning: Online Learning & BanditsMultiagent Systems: Mechanism Design
Intelligent Backhaul Link Selection for Traffic Offloading in B5G Networks
🤖 AI Summary
To address the challenges of constrained backhaul capacity, highly dynamic traffic, and stringent multi-service QoS requirements in B5G networks, this paper proposes a dynamic wireless backhaul architecture integrating network slicing, Integrated Access and Backhaul (IAB), and low-Earth-orbit (LEO) satellite communications. We introduce, for the first time, a Deep Double Q-Network (DDQN)-driven dynamic link selection mechanism tailored to IAB-satellite cooperative backhaul, enabling real-time topology reconfiguration under multi-slice QoS constraints. The agent employs a fully connected neural network with ReLU activation and converges within approximately 20 training episodes. Simulation results demonstrate that the proposed approach significantly improves backhaul resource utilization and enhances slice throughput and latency compliance rates by 23.6% and 18.4%, respectively, compared to baseline methods.
Technology Category
Application Category
📝 Abstract
Fifth Generation (5G) mobile networks considers an expansive set of heterogeneous services with stringent Quality of Service (QoS) requirements, and traffic demand with inherent spatial-temporal distribution, which places the backhaul network deployment under potential strain. In this paper, we propose to harness network slicing, Integrated Access and Backhaul (IAB) technology coupled with satellite connectivity to build a dynamic wireless backhaul network that can provide additional backhaul capacity to the base stations on demand when the wired backhaul link is temporarily out of capacity. To construct the network design, Deep Reinforcement Learning (DRL) models are used to select, for each network slice of the congested base station, an appropriate backhaul link from the pool of available IAB and satellite links that meets the QoS requirements (i.e., throughput and latency) of the slice. Simulation results show that around 20 episodes are sufficient to train a Double Deep Q-Network (DDQN) agent, with one fully-connected hidden layer and Rectified Linear Unit (ReLU) activation function, that adjusts the topology of the backhaul network.
Problem
Research questions and friction points this paper is trying to address.
B5G Networks
Intelligent Path Selection
Traffic Allocation
Innovation
Methods, ideas, or system contributions that make the work stand out.
Deep Reinforcement Learning (DDQN)
Network Slicing
Integrated Access and Backhaul (IAB)
🔎 Similar Papers
No similar papers found.
