Graph Neural Network Based Wireless Resource Management And Wireless Transmission Mechanism

With the explosion of big data and the emergence of new applications,it is expected that the future will embrace the trend of a smart society.The 6th generation(6G)communication technology needs to achieve the paradigm shift from connected people and things to connected intelligence.Artificial intelligence(AI)is regarded as a key enabler to achieve the above change.Generally speaking,implementing AI to achieve intelligent wireless networks faces three key challenges.First,how to achieve an efficient AI platform to reduce communication and computation overhead.Secondly,how to create a flexible AI platform to adapt to complex and dynamic large-scale wireless networks.Finally,how to enable a decentralized AI platform to realize the decentralized wireless network architecture.Given that graph neural networks(GNNs)can achieve efficient information exploitation,exhibit high scalability and generalization ability,and depend on naturally decentralized operations,this thesis integrates GNNs into wireless networks to deal with the above three challenges.Although GNNs are powerful for 6G,there are still some key technique challenges to fully exploit the potential of GNNs,mainly including:(1)how to design appropriate topology and node/edge features to construct graphical models for wireless communication systems;(2)how to design efficient wireless transmission mechanisms to support the decentralized implementation of GNNs in wireless communication systems.This thesis conducts in-depth research on these above problems,aiming at enhancing 6G intelligent wireless networks from the perspectives of GNN-based wireless resource management and wireless transmission mechanism.First,to address the challenge of constructing graphical models for wireless communication systems,a graphical model construction paradigm called “wireless networks as graphs” is proposed.The link scheduling problem serves as an example to validate the efficiency of this paradigm in wireless resource management.Specifically,the system model of link scheduling problem in the device-to-device(D2D)system is first introduced,based on which a wireless graphical model is designed for the D2 D network.Then,to achieve the GNN-based link scheduling,the wireless graphical model is input into the GNN and the multi-layer classifier with joint training algorithm.To further strengthen the scalability,reduce the time complexity,and solve the mismatch issue during the training stage,an unsupervised training algorithm and a K-nearest neighbor graphical model construction paradigm are developed.From the simulation results,the proposed GNN-based link scheduling algorithm can not only achieve better performance with low time complexity,but also avoid overwhelming channel state information(CSI)estimation overhead and GNN training overhead.Secondly,to strengthen the efficiency of GNNs in large-scale heterogeneous wireless networks,a graphical model construction paradigm called “wireless optimization problems as graphs” is proposed.The combinatorial auction serves as an example to validate the efficiency of this paradigm in wireless resource management.Specifically,the mathematical model of the winner determination problem(WDP)in the combinatorial auction is first introduced,based on which an augmented bipartite bid-item graph is designed for the WDP.Given that the graphical model is bipartite,the GNN-based winner determination algorithm is developed by adopting GNNs with half graph convolution and designing corresponding sample generation processes and graph-based post-processing algorithms.From the simulation results,the proposed GNN-based winner determination algorithm can reduce the time complexity of obtaining the near-optimal solution of the WDP with good scalability and generalization ability.Thirdly,to enhance the robustness of decentralized GNNs in wireless communication systems,the transmission error induced by wireless channels is studied.By exploiting the error-tolerance of GNNs,a task-oriented data retransmission mechanism is proposed for GNNs.Specifically,by theoretical analysis,the relationship between the wireless transmission error and the performance of decentralized GNN inference is built by proposing the robustness verification problem.Then,the robustness of GNNs’ prediction results in uncoded and coded wireless communication systems is discussed,where robustness metrics are defined correspondingly.Based on the above analysis,the data retransmission mechanism for GNN is proposed by re-designing novel retransmission criteria and stopping rules.The simulation results indicate that the proposed data retransmission mechanism can enhance the prediction robustness of decentralized GNN inference with less communication overhead compared with the traditional data retransmission mechanism.Finally,to protect the privacy of decentralized GNNs in wireless communication systems,the privacy leakage induced by wireless transmission is studied.By exploiting wireless channel noise,the differential privacy technique,and the over-the-air computation technique,a privacypreserving mechanism for GNN is proposed.Specifically,by analyzing the operating mechanism of over-the-air computation and decentralized GNN inference,a novel privacy-preserving wireless waveform for GNN is designed.Then,the optimal parameters of the privacy-preserving wireless waveform is obtained by solving the corresponding optimization problem.Based on the above design,the performance upper bound of decentralized GNN inference under given privacy requirement and the advantage of over-the-air computation technique in privacy preservation are analyzed theoretically.Moreover,the privacy-guaranteed training algorithm is proposed to further enhance the performance of decentralized GNN inference.Simulation results suggest that the proposed privacy-preserving mechanism for GNN works well to achieve privacy-preserving decentralized GNN inference with good performance and without unnecessary communication overhead.The research results in this thesis can provide technical solutions and a theoretical basis for GNN-based wireless resource management and wireless transmission mechanism,as well as effective ideas for the design and deployment of the next-generation intelligent wireless network.