Optimizing intelligent reflecting surfaces with discrete phase shifts and pilot overhead reduction using deep learning

Communication systems have been evolving since the first generation of communication technology, 1G. With the emergence of recent technologies, data traffic worldwide has increased, necessitating further developments in communication systems. The fifth generation of communication technology introduced base stations equipped with multiple antennas, enhancing the power and bandwidth efficiency of communication systems and leading to the emergence of certain technologies used alongside these multi-antenna base stations. Intelligent Reflecting Surfaces are expected to be a key technology for sixth-generation communication systems. Intelligent Reflecting Surfaces are a cost-effective technology that increases the energy and spectrum efficiency of communication systems. Made using numerous metamaterials, these surfaces consist of passive elements that apply a phase shift to the incoming signal. If these elements are optimized collectively, a power gain proportional to the square of the number of elements can be achieved. Despite their many advantages, when integrated into existing communication systems, IRS introduce a more complex system structure. This thesis involves configuring the phase shifts of IRS using Deep Learning. Since IRS are made up of passive elements, traditional methods of channel estimation and using this estimation to optimize phases pose challenging tasks. Moreover, traditional estimators require numerous pilot periods for channel estimation. In this study, a Deep Learning model was trained using pilot signals, and this trained model provides optimum phase values as outputs. Additionally, continuous phase shifts in practical applications make implementation difficult; thus, the model trained in this thesis provides discrete phase values. After these phase shift values are obtained, the beamforming vector at the base station is also optimized accordingly. The study examines the contribution of IRS to existing communication systems in terms of data speed, compares the performance loss relative to continuous conditions, and presents normalized mean square errors compared to the optimum scenario.