Multi-Task Generative Modeling for the Air Channel via Hierarchical GANs

Channel modeling is a fundamental problem in wireless communication. In order to obtain an accurate channel model with traditional methods, a large amount of domain expertise is required. Thus, learning-based approaches, specifically deep generative models for channel modeling have been proposed. However, deep generative modeling requires substantial amounts of training data, and while the issue of limited channel measurement data is well recognized, only few proposals to mitigate this issue exists. To fully utilise the limited data, we propose a novel generative adversarial networks (GAN) based modeling scheme for multi-task learning, inspired by Bayesian hierarchical modeling. Our proposed method leverages the similarities between similar channel generation tasks, which arise naturally as we collect measurements from different communication environments. Our simulation study shows that our proposed method successfully learns channel models from different channel datasets and outperforms the baseline solution by a large margin.