Application of Machine Learning for Radiowave Propagation Modeling Below 6 GHz

This paper presents the application of supervised learning and use of fully connected neural network (FCNN) for the development of a path specific propagation model for frequencies below 6 GHz. The model has been trained and tested against an extensive measurement dataset capturing several areas and the diverse topography of the UK. In addition to the measurement data, propagation environment related features that are provided as an input to the neural network have been extracted by employing quantised radio path profiles based on detailed topographic datasets (such as terrain, surface and land use). The results provide an insight in the choice of appropriate features that describe the propagation environment, trade-offs between prediction accuracy, extent of considered key features and associated complexity. The performance comparison between varying number of input features highlights that whilst an excellent prediction accuracy can be achieved when the training data include all test scenarios, it also reveals the limitations of pure data driven prediction methods and their inability to generalise. Noting the multifaceted nature and impact of propagation mechanisms and environments in conjunction with the limited availability of empirical measurements covering a wide range of spectrum bands, alternative hybrid approaches which may include physical modeling of propagation mechanisms are expected to further enhance the performance of neural networks based radiowave prediction methods. We present a simple hybrid approach in this paper that resolves the generalisation issues and provide far superior performance compared to pure physical propagation models or data driven models based on FCNN.