Learning Network Architecture Design for Target Localization via Mean Squared Error Bound

The architecture of a learning network is a key factor influencing its performance. Traditional learning network architecture design methods, including manual design and neural architecture search, are essentially laborious trial-anderror approaches. For the learning-based target localization, our previous work derived a closed-form mean squared error bound (MSEB), which represents the testing performance. In this work, the MSEB is used as objective function to design the learning network architecture. Compared with traditional methods, our approach obtains the learning network architecture with better performance and requires less designing time in the tested examples under the scenarios with different number of hidden layers, number of training data, maximum allowed total number of hidden layer nodes, signal-to-clutter-plus-noise ratios, and number of transmitter-receiver pairs. We summarize the arrangement of hidden layers and nodes for the optimal learning network architecture obtained by our method across various scenarios.