Utilizing Neural Networks for Fully Adaptive Radar

This paper discusses the use, of neural networks to replace the non-linear optimization at the core of the fully adaptive radar (FAR) framework. Replacement of the optimizer with the neural network reduces computational complexity, leading to faster run times. It also prepares the FAR framework for the future inclusion of learning and attention. A feedforward neural networks was trained using the Levenberg-Marquardt and Bayesian regularization algorithms along with the radial basis and generalized regression neural network architectures in order to perform parameter adaption based on the forward propagation of the prior probabilities of the target state (expected: target range, velocity and signal to noise ratio & the target range and velocity variance). The neural network was trained using data from simulated runs of the optimizer based FAR framework tracking a single target in range and Doppler. The trained neural network could approximate the optimization solutions with a lower cumulative optimization time, by an order of magnitude, and a better cumulative measurement cost. As such we concluded neural network could be used to replace the FAR framework optimizations.