Active Learning Applied to RCS Computations With Nonuniform Sampling Using Different Objective Functions

Abstract

An active learning framework is introduced to reduce the number of frequencies and angles one must consider for wideband monostatic scattering computations or measurements. This method is used to optimally select those frequencies and angles that would be most informative, resulting in nonuniform sampling and often a reduced number of points (vis-agrave-vis uniform sampling). In this paper we focus on jointly two-dimensional optimal sampling in frequency and incident angle thetas for monostatic scattering. The method consists of two basic steps. One step involves estimation of model parameters using a least-square (LS) algorithm. The next step is to optimally choose the next point (frequency and thetas) for analysis by the computational model or experiment. This new point is selected with the goal of reducing uncertainty in the parametric model (quantified via the Fisher information matrix). Iterating these two steps, a sequence of numerical computations or measurements are performed, each at the most informative point for learning the parameters of the associated simpler parametric model. This idea is demonstrated here in the context of reducing the number of points (frequencies and orientations) at which a computational model must be employed. And in order to avoid repeatedly gathering samplings at the edge of the input space, an alternative objective function is applied which makes the actively selected points closer to a region of interest