Efficient Deep Model Training for Coordinated Beam-Forming in mmWave Communications

The problem of data-driven coordinated beam-forming in a mmWave, multi-cell scenario is considered in this paper. In the relevant literature, mappings are learned via (deep) neural networks leading to considerable training minimization during normal operation, namely, after the mappings have been identified. However, the initial training of the neural networks required many data that need to be collected, which is not always feasible in the dynamically changing wireless environment. This paper studies appropriate neural network structures that exploit more effectively the underlying correlations of the involved signals in order to minimize the dataset size required for training. The efficacy of the new structures is demonstrated on the DeepMIMO dataset.