10 February 2023 in IEEE Transactions on Vehicular Technology
The prevailing issue in multi-hop wireless networking is interference management, which militates against the efficiency of traditional routing and scheduling algorithms. We develop a self-supervised learning approach to address the classic NP-hard problem of capacity optimization over a multi-hop wireless network, where the routing and scheduling decisions are deeply coupled. Our two-stage design leverages historical computation experiences to accelerate the optimization of new problem instances, where an instance represents the application-level input containing network topology, interference model, and other user-level traffic constraints. The first stage, Scheduling Structure Classification (SSC), distills the scheduling structure of the historical optimization instances into an appropriate number of classes, through a properly designed clustering algorithm without prior assumption or knowledge. The second stage uses the instances labelled with class information from the first stage to train an Application Identification (AID) neural network model capable of predicting a future problem instance's scheduling class given its application-level information. When solving the new instance, its predicted scheduling class and the associated scheduling structure are exploited to compute an efficient approximate solution, avoiding the time-consuming iterative search for such scheduling structure as practiced by the conventional approaches. We apply our method to different types of wireless multi-commodity flow problems across various network sizes and disparate flow requirements. The results demonstrate that our method significantly reduces the computation time robustly by at least 70% with only slight loss in the solution quality.
15 October 2025 in IEEE Transactions on Vehicular Technology
The ultra-densification of small cells to meet the heterogeneous traffic demands of users is a major attribute of the beyond 5G (B5G)/6G wireless networks which is still being studied. Unmanned aerial vehicle (UAV) base stations (UBSs) which can grasp tall buildings and lampposts offer flexible deployment opportunities to the B5G/6G wireless network more than do fixed base stations, as they serve as relay nodes to provide wireless backhauling for ultra-reliable low latency communications. However, the key challenge lies in jointly optimizing the UBSs deployment and multi-hop wireless backhauling to minimize the overall energy consumption. This involves finding the shortest paths with minimum energy costs, which is a mixed integer linear programming problem that is NP-hard. To tackle this problem, we propose a combinatorial optimization learning approach (COLA) that integrates a combinatorial algorithm into a neural network to quickly and effectively approximate the solution to the optimization problem. First, the neural network predicts the “magic” link weights given the application-level inputs. The link weights are then used by a simple Dijkstra's algorithm to find the shortest paths for the multi-commodity flows. We will demonstrate how to decompose the original multi-hop multi-commodity flow optimization problem into learnable single commodity shortest path problems, how to obtain the training data, and how to backpropagate the gradients incorporating the combinatorial component during training. We evaluate COLA in both small-sized and large-sized wireless networks. Our COLA greatly reduces the computation overhead, while maintaining near-optimal solution quality in terms of the energy efficiency of the suggested paths.
7 April 2026 in IEEE Transactions on Mobile Computing
The classic multi-hop wireless network optimization problem has recently re-attracted many attentions due to some emerging applications such as wireless mesh network, space-air ground integrated networks, and 5G/6G integrated access and backhaul systems. The key issue in multi-hop wireless network optimization is the interference management through scheduling. The fundamental NP-hardness of this problem is that there are exponentially many possible independent sets (ISs), but only a small number of them will be scheduled in the optimal solution (termed as the scheduling structure). The existing literature on approximation algorithms are mainly along the direction of searching the ISs in a heuristic manner. With the capability of machine learning (ML) algorithms in supporting big data analytics, in this paper, we propose a two-stage self-supervised learning framework that can explore the scheduling structure from historical optimization instances, and such knowledge is then exploited to solve a new instance with greatly reduced computational overhead. At the exploring stage, we develop dimension reduction techniques for effective scheduling structure classification in a high-dimensional vector space. At the exploiting stage, we design an innovative structure criticalness indicator (SCI)-based IS selection algorithm that can robustly lead to close-to-optimal approximation of the average achievable throughput with roughly constant complexity. Furthermore, we also contribute a geometric canonical representation (GCR) system to equip the ML-assisted optimization framework with a topology-aware applicability. The proposed topology-aware ML (TAML) framework exhibits good generalizability across different network topologies and flow demands.