A neural network approach to improve radar detector robustness

A neural network (NN) based detector is proposed for approximating the ALR detector in composite hypothesis-testing problems. The case of detecting gaussian targets with gaussian ACF and unknown one-lag correlation coefficient, ρ
s
, in AWGN is considered. After proving the dependence of the simple hypothesis-testing problem LR detector on the assumed value of ρ
s
, and the extreme complexity of the integral that involves the ALR detector, NNs are proposed as tools to approximate the ALR detector. NNs not only are capable of approximating this detector and its more robust performance with respect to ρ
s
, but the implemented approximation is expected to have lower computational cost that other numerical approximations, a very important characteristic in real-time applications. MLPs of different sizes have been trained using a quasi-Newton algorithm to minimize the cross-entropy error. Results prove that MLPS with one hidden layer with 23 neurons can implement very robust detectors for TSNR values lower than 10dB.