| الوصف: |
Resource allocation is a fundamental research topic in wireless communications. With the rapid development of wireless communication systems, the conventional optimization algorithms for resource allocation cannot meet improving requirements such as real-time execution, generalizability to larger networks, distributed implementation, and privacy. To mitigate the aforementioned challenges, this thesis focuses on developing machine learning (ML) methods to enhance resource allocation in wireless networks. Firstly, discrete optimization problems are considered. A graph neural network (GNN) based framework is proposed to solve such problems, and its effectiveness is verified by two case studies, namely link scheduling in device-to-device (D2D) wireless networks and joint channel and power allocation in D2D underlaid cellular networks. Secondly, the proposed GNN approach is extended to address a more complicated continuous optimization problem, i.e., D2D beamforming design. To reduce the learning complexity, the structure of the optimal beamforming is adopted to transfer the beamforming to primal power and dual variables. Therefore, the GNN learns these variables instead of the direct beamforming. Simulation results demonstrate that the proposed GNN approach achieves better sum rate than the state-of-the-art benchmarks, and reduces the execution time to millisecond-level which is favorable in the time-stringent applications. Besides, the GNN approach has great features desired in wireless communications including the generalizability to large networks without retraining and the robustness to corrupted input features. However, the implementation of the proposed GNN method is centralized. Consequently, a decentralized federated learning (FL) approach is proposed to address the D2D link scheduling problem, in which a shared model is distributedly trained by each client, hence the privacy of each client is guaranteed. This is the first work that utilizes FL to solve the D2D link scheduling problem in a distributed manner. Numerical results show that the proposed FL method achieves almost the same performance in terms of the accuracy and the sum rate as the centralized training. |