Weakly-Supervised Generative Adversarial Nets with Auxiliary Information for Wireless Coverage Estimation.

Wireless coverage is the received signal strength of a particular region, which is a key prerequisite to provide high quality mobile communication service. In this paper, we aim to estimate the wireless coverage of an area based on the randomly distributed samples of received signal strength collected within the area. Specifically, we propose a weakly-supervised generative adversarial nets with auxiliary information (WS-GAN), which is a data-driven model to make estimation on the wireless coverage. Different from conventional methods (e.g., k-NN) that predict the missing values by simply replacing them with the average of nearby observations, WS-GAN adopts a GAN-based structure, in which the generator approximates the real distribution of observations under the supervision of the discriminator, while the discriminator tries to distinguish the ground truth from "faked" data produced by the generator. WS-GAN differs from the literature in two aspects: (1) WS-GAN is able to use the auxiliary information of local radio environment, e.g., geographical information as terrain and buildings, to improve the estimation performance, since such auxiliary information has significant impact on the variation of received signal strength. (2) The objective function of WS-GAN combines adversarial loss and a new light-weight reconstruction error term for weakly supervision, which are jointly optimized during training. The experiments on a real dataset show that WS-GAN can generally achieve more accurate results than baselines. Moreover, through a case study, we demonstrate that the wireless coverage maps generated by WS-GAN are more rational and practical than those obtained by baselines.