Machine Learning Accelerated Antenna Selection for an Underlay Spectrum Sharing System

Machine learning (ML) techniques can help reduce the computational complexity of antenna selection (AS) for an underlay spectrum sharing system, which enables efficient spectrum utilization. We illustrate this by proposing two ML approaches: i) A supervised learning based classification approach with instantaneous channel power gains as input and antenna index as output. It predicts antenna index directly with low complexity; ii) A model-based regression approach with system parameters such as the number of antennas, fixed transmit power, and channel statistics as inputs and a parameter required to implement the conventional AS rule as output. It substitutes the obtained parameter in the conventional $A S$ rule to select the antenna index and exploits the structure of the optimal AS rule. Furthermore, it generalizes to different system parameters and reduces training overhead significantly. We show that both neural network and random forest models trained for the above approaches are near-optimal. We also show that these models work for different channel models including 3GPP adapted tap delay line models while requiring orders of magnitude lower amount of time.