Conventional multi-user scheduling in cellular networks are required to make a decision every transmission time interval (TTI) of at most several milliseconds. Only quite simple schemes can be implemented under the stringent time constraint, resulting in far-from-optimum performance. In this paper, we focus on the case of scheduling multiple users in a fixed wireless access (FWA) network with stable channel characteristics. We propose a scheduling approach by which a high-quality scheduling decision based on statistical channel state information (CSI) is made across all TTIs instead of making simple TTI-level decisions. The proposed design is essentially a mixed- integer non-smooth non-convex stochastic problem. We first replace the indicator functions in the formulation by smooth sigmoid functions to tackle nonsmoothness. By leveraging deterministic equivalents (D.E.), we then convert the original stochastic problem into an approximated deterministic one, followed by linear relaxation of the integer constraints. However, the converted problem is still nonconvex due to implicit equation constraints formerly introduced by D.E. Therefore, we employ implicit optimization technique to compute the gradient explicitly, with which we further propose an algorithm design based on a modified version of Frank-Wolfe method. Numerical results verify the effectiveness of our proposed scheme.

Network optimization plays a crucial role in wireless communications. However, the optimization of contemporary 5G networks is challenging due to its black-box nature and huge searching space. To address these challenges, this paper introduces efficient zeroth-order (ZO) algorithms and a parallel framework for optimizing large-scale networks. By leveraging the gradient-based searching strategy, the proposed algorithms, namely ZO projected gradient descent (ZO-PGD) and ZO block coordinate projected gradient descent (ZO-BCPGD), can significantly improve optimization quality and computational efficiency. Both algorithms guarantee a convergence towards stationary points under mild conditions, eliminating inherent errors in traditional ZO methods. We further propose a parallel framework for optimizing the network parameters in a simultaneous manner. By partitioning the entire network into manageable subnetworks, the original network optimization problem is reformulated as a consensus optimization problem and tackled in parallel using the penalty dual decomposition (PDD) method. We have also designed a tailored size-constrained grid clustering algorithm for network partitioning to ensure load balance among parallel working nodes. The efficacy of our proposed schemes is verified by extensive numerical results. Our ZO algorithms outperform existing methods in both computational efficiency and solution quality. Moreover, the parallel framework significantly reduces execution time without sacrificing network performance.