Constrained Online Learning to Mitigate Distortion Effects in Pulse-Agile Cognitive Radar

Pulse-agile radar systems have demonstrated favorable performance in dynamic electromagnetic scenarios. However, the use of non-identical waveforms within a radar's coherent processing interval may lead to harmful distortion effects when pulse-Doppler processing is used. This paper presents an online learning framework to optimize detection performance while mitigating harmful sidelobe levels. The radar waveform selection process is formulated as a linear contextual bandit problem, within which waveform adaptations which exceed a tolerable level of expected distortion are eliminated. The constrained online learning approach is effective and computationally feasible, evidenced by simulations in a radar-communication coexistence scenario and in the presence of intentional adaptive jamming. This approach is applied to both stochastic and adversarial contextual bandit learning models and the detection performance in dynamic scenarios is evaluated.