LEARNING NETWORKS FOR EXTRAPOLATION AND RADAR TARGET IDENTIFICATION

The problem of extrapolation for near-perfect reconstruction and target identification from partial frequency response data by neural networks is discussed. Because of ill-posedness, the problem has traditionally been treated with regularization methods. The relationship between regularization and the role of hidden neurons in layered neural networks is examined, and a layered nonlinear adaptive neural network for performing extrapolations and reconstructions with excellent robustness is set up. The results are then extended to neuromorphic target identification from a single "look " (single broad-band radar echo). A novel approach for achieving 100% correct identification in a learning net with excellent robustness employing realistic experimental data is also given. The findings reported could potentially obviate the need to form radar images in order to identify targets and could furnish a viable and economical means for identifying noncooperative targets.