Self-learning Constellation Mapping Method Based on Neural Networks
Yingzhe Luo, Jianhao Hu · 2020
Neural networks have been gained extensive study and application in the communication systems. This paper applies the neural networks to the transmitter, which proposes a self-learning constellation mapping method. The proposed method contains two kinds of neural networks. One is used to generate a new constellation mapping which can compensate the channel distortion. The other one is used to determine whether the newly generated constellation mapping is optimal. These two types of neural networks constitute an actor-critic architecture which is trained through the reinforcement learning method. Without losing generality, this paper applies the proposed method to the quadrature amplitude modulation systems. In the simulation experiments, the proposed method can generate some new quadrature amplitude modulation constellation diagrams which can compensate in advance the amplitude and phase distortion caused by the channel. Moreover, the performance of the proposed method is superior to the classical compensation method. Significantly, the neural networks in the proposed method are trained online, so there is no need to prepare a large amount of artificial data sets in the training phase.