DNN assisted Sphere Decoder

Abstract

A modified sphere decoding (SD) scheme is proposed for multiple-input multiple-output (MIMO) communication systems in this paper. The contribution of the paper includes the introduction of a systematic approach to sphere radius design and control based on Deep Neural Networks (DNNs) as well as the complexity advantage yielded by the proposed scheme. The learning model is introduced to predict the number of lattice points inside the sphere with some radius. Since this number is cleverly learnt by a neural network (NNW), the SD updates the radius until expecting a small number of points and then starts the search hypersphere, which greatly reduces the computational complexity. We show through simulation that for high dimensional MIMO systems the number of lattice points highly reduces in the new SD algorithm, which leads to a complexity only 3 times of the MMSE decoder complexity.