Reinforcement Learning Algorithm for Improving Spectral Energy Efficiency Using Large Intelligent Surfaces

ABSTRACT The Spectral Energy Efficiency ( SEE ) is the concrete feature of future generations of wireless systems. It is in turn dependent upon the System User‐Achievable‐Data Rate ( SAR ). The SAR of the current generation systems can be enhanced by use of Large Intelligent Surfaces (LIS). They implement a pane of reflecting antennas made up of meta‐materials. These panels are mounted on any architectural structure like apartments, schools/colleges etc. The beauty of LIS is that they can be trained by means of machine learning models to reflect the incoming electro‐magnetic signal towards the required direction that can increase the received signal strength at the receiver. This increased signal strength at the receiver further boosts the Signal to Noise ratio ( SNR ) and SAR. This paper implements a Reinforcement Learning (RiL) based customized loss model in a Recurrent Neural Network (RNN) model to enhance the SEE of the LIS based systems. The dataset required for training and validation of DL model is produced from the publicly available ray tracing based DeepMIMO generator. The simulation findings demonstrate that the suggested RNN‐RiL model exhibits an enhancement of 1.14 bps/Hz in SAR , and an improvement of 2.75 Mbits/J enhancement in the SEE when compared to the baseline technique. This rise in the SEE can be useful in inculcating more number of users per sec while maintaining the Quality of Service (QoS) thus enabling energy harvesting in LIS.