Existing intelligent jamming decision-making methods commonly rely on the perfect information assumption of the radar working state, which is difficult to establish in actual non-cooperative adversarial scenarios. To address this issue, this paper proposes a joint optimization method for jamming types and power based on hierarchical deep reinforcement learning (HDRL), aiming to achieve efficient jamming decision-making and adaptive resource allocation under uncertain environments. First, the sequential decision-making problem of optimizing discrete jamming types and continuous power parameters is modeled as a partially observable Markov decision process (POMDP), effectively characterizing state uncertainty and hybrid action space properties. Second, the HDRL architecture is designed: the upper layer addresses observational uncertainty by constructing a radar belief state and employs a Prioritized Experience Replay Dueling Deep Q Network (PER-Dueling DQN) to make jamming-type decisions; the lower layer, informed by the jamming type selection results, utilizes the Post-Decision Proximal Policy Optimization (PDPPO) algorithm to precisely control power parameters. By designing a comprehensive reward function, the agent is guided to collaboratively learn effective jamming types and their corresponding power ranges. To enhance the practicality and generalization ability of the strategy, a policy distillation technique is further introduced to integrate adversarial sample experiences under different radar state transition rules and power constraints. Research indicates that the proposed method can achieve efficient adaptive allocation of jamming resources under uncertain radar state observations, providing a new solution for intelligent jamming decision-making in complex electromagnetic confrontation environments.
Li et al. (Sun,) studied this question.