In comparison to traditional radar countermeasure system, the closed-loop behavior learning is introduced into Cognitive Radar Countermeasure(CRCM), which conducts state recognition and jamming effect evaluation through radar signals, and then jamming strategy is optimized by autonomy making jamming more initiative and pertinent. Radar state recognition is the basis of CRCM, but the target radar may activate previously hidden unknown states in the opposed process, which compels CRCM to react to the unknown states rapidly. In consequence, unknown radar state recognition is focused on for CRCM, and two recognition methods are proposed based on supervised classification and unsupervised clustering respectively utilizing related machine learning algorithms. The simulation results validate the effectiveness of the two approaches.