Neural Network With Binary Cross Entropy for Antenna Selection in Massive MIMO Systems: Convolutional Neural Network Versus Fully Connected Network

This paper makes several contributions to antenna selection techniques for massive multiple-input multiple-output (mMIMO) systems using artificial neural networks. First, binary cross entropy is adopted as the loss function for network training instead of the conventional cross entropy, which reduces the number of nodes in the output layer from
$\binom {N_{R}}{N_{RS}}$
to
$N_{R}$
, where
$N_{R}$
and
$N_{RS}$
are the number of candidate antennas and the number of selected antennas, respectively. In mMIMO systems, which have a large number of antennas, binary cross entropy is essential. We also demonstrate that the channel matrix is practically sufficient information to train the network, excluding the signal-to-noise ratio (SNR) factor present in the capacity formula. Since a single label is generated for a given mMIMO channel regardless of SNR, the size of training data is reduced significantly. When the channel matrix without pre-processing is inputted into a neural network for feature extraction, which is referred to as pure connectionist feature extraction, we show that the convolutional neural network (CNN) extracts features more successfully than the fully connected network (FCN). We also show that hybrid feature extraction, in which features are first extracted symbolically from the channel matrix and then connectionist features are extracted from the symbolic features, offers significant performance improvement over pure connectionist feature extraction from the raw data. However, when features are extracted in a hybrid manner, FCN achieves marginally better performance than CNN, contrary to the pure connectionist feature extraction. Finally, when the networks in the hybrid feature extraction are pruned to be suitable for deployment in mobile devices, we show that FCN is a better choice, as it is more robust to severe pruning than CNN. We conducted computer simulations to demonstrate the effectiveness of the proposed approaches.