A Fast-Learning Sparse Antenna Array

Selecting a sparse subset of antennas to obtain high-resolution direction-of-arrival estimates while circumventing the complexity associated with using a large array is critical in many radar applications. Since this subset selection problem is combinatorial, deep learning has been recently proposed as a possible solution for efficiently solving it. However, the bottleneck in this approach is training data generation, which requires an exhaustive search over all possible subarrays. In this paper, we propose an efficient method for generating training data using ideas from submodular optimization. In particular, we use the log-determinant of the Cramér-Rao lower bound as our cost function due to its submodular structure. It is then minimized through a greedy optimization approach to determine the best subarray. We provide numerical simulations to validate the performance of the proposed array selection strategy. Our simulations show that the proposed approach is ten times faster in training than an exhaustive search method while providing comparable performance.