Research On Message Passing Algorithms Assisted By Neural Networks

In recent years,the relevant technologies of deep learning have been introduced into multiple key areas of mobile communication,such as large-scale MIMO,LDPC,and NOMA.Message passing algorithm,as a common iterative algorithm in communication systems,can be implemented in the form of factor graphs for intuitive representation.Therefore,based on the structural similarity between factor graphs and neural networks,this paper proposes a neural network-assisted message passing algorithm.With the goal of optimizing the message passing algorithm using neural networks,this paper first conducts research on the optimization of decoding algorithms under neural network-assisted Rayleigh fading channels.By introducing learnable weights in the factor graph to weaken the influence of short loops on correlation,the decoding performance is improved.Based on this,a deep investigation is conducted on the distortion of channel state information at the receiver,classifying channel conditions according to the degree of distortion,and obtaining decoders with complexity and performance balance respectively.The paper uses the extrinsic information transfer chart(T-EXIT)as a theoretical analysis tool to evaluate the performance of decoders,guide the selection of training sets,and determine the approximate scheme for the channel state information.Finally,the idea of optimizing message passing algorithms in factor graphs based on neural networks is extended to multi-user detection algorithms.According to the differences between decoding algorithms and detection algorithms,the proposed neural network detector is improved,achieving significant performance improvement under single-user channel state distortion.The simulation results show that the proposed neural networkassisted message passing algorithm has achieved good performance in both decoding and detection algorithms.Additionally,using the T-EXIT theory as a guiding tool for decoding scheme selection has practical significance.