Power Coverage Optimization in Cellular Radio Networks Using Neural Architecture Search

Due to the complex propagation environment, it is difficult to find the optimal antenna configuration that ensures the best coverage of cellular networks. Most existing work adopts heuristic methods, where the antenna configuration is evaluated and updated by calculating the coverage of the entire area iteratively. However, in real networks with a large number of antennas and highly complex ray-tracing models, the heavy computational burden caused by repetitive coverage calculation is usually unacceptable. In this paper, we propose a deep learning approach that straightly infers the antenna configuration from the raw data given by the city map. Specifically, we adopt the emergent neural architecture search to automatically design a specific neural network for the considered power coverage optimization problem. Numerical results show that the discovered neural architecture greatly outperforms the conventional hand-crafted architectures.