Nonlinear Waveform Sensing for Cognitive Radar Based on Reinforcement Learning

Cognitive radar automatically adjusts its waveform via ceaseless interaction with the environment and learning from the experience. Compared with the linear frequency modulation (LFM) that has been commonly adopted in cognitive radars, the nonlinear FM (NLFM) signal has more flexible frequency variation and small time delay–Doppler coupling. In this work, we propose an NLFM cognitive radar based on reinforcement learning for target sensing. Specifically, a radar waveform selection framework is proposed via the interactive multimodel. It embraces the Riccati equation and Riccati-like iterative calculations to obtain the prediction error covariance (PEC) and the prediction Bayesian Cramér–Rao lower bound (PBCRLB), respectively, which are used to guide the optimal waveform design. With PEC or PBCRLB, an entropy reward Q-learning method is also proposed for joint waveform parameter selection (JWPS) and pure waveform parameter selection from the NLFM library. Simulations show that both the time complexity and tracking accuracy of PEC-based Q-learning JWPS outperform that of the PBCRLB method. Furthermore, PXIe-5785 is utilized to construct a cognitive radar platform and conduct field experiments for nonlinear waveform sensing, which confirms that nonlinear waveforms are more effective than linear waveforms in target localization.