Deep Learning methods have shown that they can be efficiently used to estimate radar parameters on synthetic datasets that were used both for training and performance evaluation1 2. However, applying these methods to real measurement data remains challenging. This is mostly due to the scarcity of labeled measurement data and the associated difficulty of obtaining sufficiently labeled datasets3.
For first and second-order optimization methods, one usually is in need of two things. A step direction and a suitable stepsize to allow rapid convergence. Usually one employs a condition in the form of 1 or uses a so-called Trust-region2 to limit the stepsize during the iteration. If we look at it from another standpoint, we can consider for example a set of various different step directions and we want to find a suitable weighting of those in order to minimize the cost-function as rapidly as possible. Finding these weighting factors is the job of a step direction selection method.
