Robust Intrapulse Frequency Coding Against Time-Gated Frequency-Following Jamming via Belief-State Reinforcement Learning

Constrained by instantaneous receiver bandwidth and intermittent jamming mechanisms, radar antijamming against time-gated frequency-following jammers inherently degrades into a time-frequency dual-domain partially observable Markov decision process. To overcome these physical constraints, this article proposes an end-to-end belief-state reinforcement learning framework for robust intrapulse frequency coding. By directly mapping short-time Fourier transform sequences to agility commands, the architecture eliminates feature truncation errors. Furthermore, a state-transition-aware reward mechanism is designed to establish a phase transition boundary, analytically resolving the exploration-exploitation dilemma under partial observability. Feature manifold analysis reveals that the agent dynamically reconstructs the jammer's generative mechanism rather than memorizing static sequences. Simulations demonstrate that the method achieves near-optimal steady-state robustness and enables rapid policy reconvergence against nonstationary strategy shifts without offline retraining.