Application of Learning-based Resource Allocation Scheme for Different UE Antenna Orientations

Abstract: One of the main tasks of the fifth generation (5G) network systems is the ability to provide high-data rates with always-on connectivity to the ever-increasing number of smart devices, especially for high mobility users. By cooperation within a number of the Remote Radio Heads (RRHs), higher system capacity can be achieved based on the Cloud Radio Access Network (CRAN) architecture for the 5G network systems. For the baseband processing, the state of the art technology uses the instantaneous channel state information (CSI), which will be infeasible for higher mobility users; fi rst because it will incur signifi cant system overhead, which will reduce the effective system performance, and second because infrequent CSI acquisition will get outdated for such high mobility users. In contrast to this method, a novel approach is proposed to use the users' position information for baseband processing (or resource allocation) to achieve better system performance. For this purpose, the machine learning algorithm, named "random forests", is used to learn the correlation of users' position information with the different system parameters. The proposed scheme only requires the users' positions estimates so that the overhead problem, which means a degradation in the spectrum utilization, can be avoided. The random forests is used as a multi-class classi fier and it achieves 92% of the system goodput compared to the traditional CSI-based scheme. In the thesis work, the novel approach is applied in a more realistic scenario where multiple smart devices have an arbitrary antenna orientation, unknown to the RRH. In this case, the system goodput from the learning-based resource scheme reduces to about only 27% of the goodput from traditional CSI-based scheme. This thesis provides two problem fi xes in such mismatch cases for real-time User Equipment (UE) antenna orientation. One method is based on rotating the predicted receive fi lter direction according to the actual user-antenna orientation, which shows the system goodput improvement to 86% of the CSI-based scheme. The other method uses the user-antenna orientation as an input feature when training the random forests classifi er, which makes the system goodput to be about 95% of the goodput for CSI-based scheme.