Path loss and Interference modelling of mmWave networks using Machine learning

End-to-end network performance evaluation and dynamic resource provisioning require models that are fast in execution and produce accurate predictions. 5G mmWave mobile networks are challenging for developing such models due to the difference in line of sight (LoS) and non-line of sight (NLoS) regimes. The thesis explores the possibility of using data driven Machine learning to develop such models for mmWave networks. It consider two probabilistic models for path loss prediction and a neural network (NN) model for SINR prediction in mmWave networks. For path loss, a Bayesian learning and a Mixture Density neural Network (MDN) model is developed and trained to predict path loss distributions in a realistic city environment based on a limited amount of training data. The ability of the models to predict distribution allowed to also take into account the uncertainty in the predicted value. The prediction capability of the models is measured by comparing true path loss distribution with the predicted distribution. Kullback-Leibler (KL) divergence and Total Variation Distance (TVD) are used for comparing the distribution. The results show that MDN describes path loss more accurately for larger training data-sets. However, a Bayesian learning predictor is more data-efficient. Similarly, for SINR a NN is trained to predict SINR traces. In addition to KL and TVD, the SINR model is also evaluated in terms of absolute error. The evaluation shows that the SINR model provides moderate prediction accuracy, leaving room for further refinement. The thesis also hints at some of the approaches that can be explored in future works