End-to-end Physical Layer Optimization Scheme Based on Deep Learning Autoencoder

5G large-scale commercial enterprises will provide high-speed and low-latency communication services, but it also challenges optimization of system. In particular, the conventional block-by-block design communication model has less interpretability and increases the complexity of physical layer optimization. Fortunately, deep learning has an inherent advantage in structured information representation and data extraction. Therefore, in this paper, based on deep learning autoencoder scheme is proposed to perform end-to-end physical layer optimization. Firstly, a two-stage training mode is proposed to improve the generalization of the neural network. Secondly, in order to decrease the system overhead caused by channel state information (CSI) feedback, utilizing autoencoder to reconstruct CSI by its compression feature. The simulation results show that the phased training can effectively improve the convergence rate, the compression and quantization of CSI alleviates the system loads.