GC-DRQN: Enhancing Radar Anti-Jamming Performance With Supervised Auxiliary Tasks and Deterministic Rewards

Self-defense suppression jammers pose a critical threat to radar by adaptively altering jamming frequencies based on intercepted radar pulses, which can mask the real targets. An effective countermeasure is to transmit a cover pulse before the detection pulse to deceive the jammer. In order to maximize radar detection performance while ensuring successful anti-jamming, reinforcement learning (RL) methods are employed to dynamically adjust the width ratio between the cover pulse and the detection pulse based on the jamming state and reward feedback. However, unknown jammer interception durations and interception-jamming cycles, along with the random feedback affected by noise, pose significant challenges to the pulse width selection based on RL methods. Inspired by biological intelligence to enhance RL, we propose a supervised learning (SL)-based general auxiliary task framework that emulates the spatiotemporal encoding characteristics of grid cells and time cells in mammalian brains to extract richer and more structured environmental information. Building on this, we introduce a flexible grid cell-deep recurrent Q-network (GC-DRQN) architecture, integrating SL and RL, which improves the performance of RL in handling tasks with temporal dependencies. Additionally, we implement a deterministic equivalent reward mechanism to overcome the instability in the RL convergence process caused by random rewards. Simulation results demonstrate that the pulse transmission strategy learned by GC-DRQN achieves significantly higher target detection probabilities compared to several baseline methods. Notably, in a low signal-to-noise ratio (SNR) scenario, GC-DRQN improves the target detection probability and convergence speed of DRQN by up to twofold.