Optimizing LIS-assisted wireless communication systems using deep learning

In this thesis, we propose a comprehensive deep learning-based beamforming for double-sided Large Intelligent Surface (LIS) communication in the next-generation computer communication and intelligent wireless network systems. The method uses a multi-layer perceptron (MLP) to predict the optimal reflection beamforming vector using channel state information (CSI)-related environment description. Consequently, the system is able to dynamically adapt to channel variations with no need for iterative optimization and hence can perform efficiently in real-time applications. The performance of the proposed model is evaluated based on four benchmark datasets, including DeepMIMO, MIMO Radar Signal Dataset, Real-Time MIMO IoT Antenna Selection Dataset, and WiFi CSI Localization Dataset. The results confirm the better performance of Adam in terms of Bit Error Rate (BER) minimization with up to 25% reduction across the datasets. Furthermore, the achievable rate on the DeepMIMO dataset improved from 8.37 bits/s/Hz under RMSPROP to 8.50 bits/s/Hz with Adam. This gain, and also the improved supportability against practical system constraints, result in increasing the spectral efficiency and improving device performance further under realistic limitations including reduced active elements, constrained training power, and limited training overhead.