Channel Beam Pattern Extension for Massive MIMO via Deep Gaussian Process Regression

For a typical small-sample challenge in the antenna optimization for the fifth generation (5G) massive multiple-input multiple-output (MIMO) systems, i.e., the required channel beamspace statistics or channel beam patterns are insufficient in the dimension of user position and base station (BS) array orientation, an extension method of channel beam pattern is proposed based on neural network (NN) driven Gaussian process regression (GPR). The proposed scheme utilizes the NN to transform the user position and BS array orientation into features more suitable for the basic kernel function, thus reducing its model error. Meanwhile, the small number of hyperparameters in the GPR model reduces the possibility of overfitting compared to deep NN (DNN). Simulations based on the open Deep MIMO dataset validate the superiority of the proposed scheme in channel beam pattern extension compared to the basic GPR and DNN based methods.