A Unified Framework Combining Feature Extraction and Reward Estimation for Cognitive Radar Policy Prediction

The increasingly flexible decision-making process of modern cognitive radar (CR) systems has placed considerable challenges toward electronic reconnaissance systems or radar warning receivers, which must help to determine the appropriate countermeasure methods for these radar systems in modern complex electromagnetic environments. It is necessary to accurately predict CR's internal action policy according to the interaction observations between CR and receivers under non-cooperative scenarios. Such that useful adversarial information can be extracted by reconnaissance systems for afterward processing and analysis. From the perspective of the reconnaissance side, this work presents a unified CR action policy prediction framework that combines feature extraction and reward estimation processes without much prior knowledge about CR's decision-making process in terms of CR reward structure and components. Focusing on the CR performing continuous control tasks, a method based on Bayesian nonparametric (BNP) theory is designed for automatic feature extraction from observed demonstration trajectories that are collected in various CR tasks. The proposed BNP method can significantly improve the multi-dimensional information representation ability of features. Further, a deep inverse reinforcement learning (DIRL) method is designed to estimate CR's internal reward functions without the knowledge of function structures. The performance of the proposed method is evaluated and verified under simulated CR target tracking scenarios. Experimental results showed the superiority and effectiveness of the proposed method.