Improving cellular communication system performance by pattern learning

Motivated by the fact that the achievable quality of service in cellular mobile communications depends on the multi-user channel capacity and its time-varying nature, this thesis targets the application of pattern learning in order to enhance cellular network performance. Patterns are specific features of the environment or users that appear repeatedly and have impact on optimization problems in link-level, single-cell and distributed multi-cell protocol decisions. The approach consists of determining the patterns to be learned and developing algorithms to solve optimization problems. Three types of patterns are considered in this work: 1) features of the environment, e.g., channel characteristics with geographical information (channel gain and correlation); 2) user behaviors, e.g., user mobility history in terms of past trajectories; 3) optimized parameters and solutions in the past. The patterns are stored and updated periodically by collecting constant measurements and optimized solutions. This thesis investigates the achievable improvements based on the learned patterns in different scenarios. On the link level, adaptive channel estimation and prediction are enhanced based on a-priori information about channel frequency-spatial characteristics. Furthermore, a quantized feedback scheduling algorithm is proposed, that exploits information about Doppler shifts and channel conditions. On the cell level, the improved performance is realized by long-term window resource allocation based on future channel conditions, which are determined by the location-dependent channel gain pattern, together with user location prediction. On the network level (multi-cell scenario), long-term resource allocation is on one hand complemented with handover prediction, according to scheduling window length and user mobility. On the other hand, it is complemented with enhanced inter-cell interference coordination (eICIC) in heterogeneous networks. To reduce the computational complexity of the proposed algorithms and allow for implementation on resource constrained hardware, the approximation functions for the eICIC parameter mappings are determined by machine learning. In a quite different application, it is studied how far the knowledge about location-dependent channel statistics can be employed in physical layer security, where the data rate of an eavesdropper at an unknown location is bounded by additional deliberate transmission of directional co-channel interference