摘要 : Deep learning enables real-time resource allocation for ultra-reliable and low-latency communications (URLLC), one of the major use cases in the next-generation cellular networks. Yet the high training complexity and weak generali... 展开 Deep learning enables real-time resource allocation for ultra-reliable and low-latency communications (URLLC), one of the major use cases in the next-generation cellular networks. Yet the high training complexity and weak generalization ability of neural networks impede the practical use of the learning-based methods in dynamic wireless environments. To overcome these obstacles, we propose a parameter generation network (PGN) to efficiently learn bandwidth and power allocation policies in URLLC. The PGN consists of two types of fully-connected neural networks (FNNs). One is a policy network, which is used to learn a resource allocation policy or a Lagrangian multiplier function. The other type of FNNs are hypernetworks, which are designed to learn the weight matrices and bias vectors of the policy network. Only the hypernetworks require training. Using the well-trained hypernetworks, the policy network is generated through forward propagation in the test phase. By introducing a simple data processing, the hypernetworks can well learn the weight matrices and bias vectors by inputting their indices, resulting in low training cost. Simulation results demonstrate that the learned bandwidth and power allocation policies by the PGNs perform very close to a numerical algorithm. Moreover, the PGNs can be well generalized to the number of users and wireless channels, and are with significantly lower memory costs, fewer training samples, and shorter training time than the traditional learning-based methods. 收起