Decision-Aware Waveform Design With POMDP Selection for Sequential Classification by HRRPs
Marcel Warnke, Kilian Barth, Joachim Ender · IEEE Transactions on Radar Systems · 2026
A cognitive radar benefits from adapting its transmit waveforms to the current sensing task and environment.This paper presents a framework for designing waveform libraries that improve sequential target classification using highrange resolution profiles (HRRPs). The closed-set classificationis performed with a correlation-based classifier operating onnormalised HRRPs, and the sequential selection of waveformsis formulated as a partially observable Markov decision process(POMDP), where the actions correspond to transmit waveformsor terminal class decisions and the observations are obtainedfrom the classifier outputs. Given simulated or measured targetimpulse responses, we use a genetic algorithm (GA) to synthesise constant-envelope nonlinear frequency-modulated (NLFM)waveforms that influence the classifier’s performance. Two designstrategies are considered: (i) a metrics-driven library, whereeach waveform maximises sensitivity, precision, or F1-score fora specific class, and (ii) a decision-aware library, where thefitness of a waveform is the maximum sensing cost at whichthe POMDP still selects it, directly quantifying its cost–benefitwithin the sequential decision process. Simulations with threeand four target classes show that both libraries reduce therequired number of measurements and increase the probability ofcorrect classification up to 20.28 % compared to a fusion of linearfrequency modulations (LFMs). Measurements with miniaturisedcar models are conducted to confirm the performance gains anddemonstrate robustness to moderate mismatch in aspect angleand prior information.