Active Learning for Radar System Performance Verification

This paper studies active learning methods for the verification of radar systems. Verification is based on evaluating the system for varying parameters-observation pairs. To reduce the number of observations, an evaluation-learning-selection cycle is introduced for radar systems aimed at reducing the number of samples by on-the-fly sampling. Emphasis is given to define a framework in which various strategies can be used to select the next-sample on-the-fly. The framework subdivides the sampling domain into subdomains to select the next-sample, gauging discrepancy levels within these areas to guide the selection of subsequent parameters. Additionally, we proposed a set of uncertainty-quantifier functions for the various regression methods employed in the learning stage. By comparing these methods using a radar detection performance example, the competitiveness of cost-effective approaches in adaptive sampling for the verification of the radar system performance is illustrated.