Neural Network-Based Control of an Adaptive Radar

This article demonstrates the use of a neural network to replace the nonlinear optimization at the core of a fully adaptive radar (FAR) system. Use of the neural network is shown to reduce computational complexity, leading to faster updates of the radar parameters. The feed-forward neural networks used were trained using the Levenberg–Marquardt algorithm along with a generalized regression neural network architecture. Training data was obtained from an existing FAR optimization system using the system’s parameter selections and target states for single-target tracking in range and velocity. The trained networks were then used to perform radar parameter selection based on the predicted (forward propagation of prior probabilities) probabilities represented by the target state (target range, velocity, and signal-to-noise ratio) and error variance (target range and velocity variance). The trained neural network was able to reduce the average optimization time by an order of magnitude while obtaining on average a lower resource consumption and similar radar task performance.