Stochastic Complex-valued Neural Networks for Radar

Neural networks (NNs) prove to be performant in learning nonlinear models, but their mechanisms are yet to be fully understood. Since signal models in radar are inherently nonlinear with respect to unknown range, Doppler or angles, and moreover, radar processing is intrinsically stochastic, stochastic NNs which tie the numerical capability of NNs with the probabilistic inferences can enhance model-based radar processing. Indeed, radar data are complex-valued while most algorithms based on NNs are real-valued and furthermore, lack of uncertainty assessment. To address these issues, we elaborate, in the present paper, a stochastic complex-valued NNs framework for radar. We show that these networks can achieve parameter estimation with refined learned models from radar measurements and provide an indicator of the uncertainty on the estimation. We also build a stopping criterion based on the detection principles, so that the NNs training stops when there is noise only in data. Finally, the performances of the networks are illustrated in simulation.