@article{WU2025, 
author = {Zhenhua WU and Jun QIAN and Lei ZHANG and Lili CHEN and Ning ZHU and Lixia YANG},
title = {Designing online anti-jamming strategy for agile frequency radar via Thompson sampling},
year = {2025},
journal = {Journal of National University of Defense Technology},
volume = {47},
number = {5},
pages = {206-215},
keywords = {radar frequency agility, online learning, multi-armed bandits, Thompson sampling},
url = {https://www.sciopen.com/article/10.11887/j.issn.1001-2486.23060016},
doi = {10.11887/j.issn.1001-2486.23060016},
abstract = {In the context of dynamic countermeasures between radar and active jammer, the working frequency of radar and adversarial jammer were modeled as the combat action space based on the multi-arm bandit (MAB) model in online learning theory. By exploring the uncertainty of the jamming environment state through multiple-round pulse transmission, a frequency channel jamming recognizer based on a convolutional neural network was constructed to obtain the posterior probability estimation of the belief state of each frequency channel. Then the Thompson sampling algorithm efficiently solved the built MAB model, achieving a balance between exploration and exploitation. Simulation results show that compared with random frequency agility and deep reinforcement learning algorithms, the method had higher convergence performance and was more adaptable to dynamic fast-changing jamming environments, which can give full potential to the antagonism advantage of radar active waveform transmission.ObjectiveIn the context of modern informational warfare, where electromagnetic spectrum dominance is critical, radar systems face significant challenges from complex and dynamic jamming environments. Existing radar frequency agility strategies (such as manual frequency hopping, adaptive frequency agility, etc.) often rely on human operation and pre-defined rules, lacking real-time closed-loop utilization of jamming environment information. This makes them inadequate for rapidly adapting to the real-time, fast-changing, and non-cooperative nature of adversarial jammer tactics, consequently limiting the speed of the "Observe-Orient-Decide-Act" (OODA) loop. Therefore, there is an urgent need to develop a novel approach capable of real-time perception of the dynamic jamming environment, efficient identification of jamming states, and autonomous, intelligent adjustment of radar transmission frequency strategies. This aims to fully leverage the radar's inherent advantage as the active emitter, enhancing its counter-jamming effectiveness and survivability in complex, rapidly evolving interference scenarios. Addressing this key problem is the primary objective of this research, holding significant theoretical and military importance.MethodsThis paper proposes a radar anti-jamming frequency selection strategy based on an online learning theoretical framework. Specifically, the dynamic frequency-domain confrontation between the radar and jammer is formulated as a MAB model, where each available operational frequency channel is treated as an "arm". The radar explores different frequency channels through multiple rounds of pulse waveform transmission, collecting echo signals. To address the uncertainty in the jamming environment state (i.e., the inability to directly observe the jammer's exact status), a convolutional neural network (CNN)-based frequency channel jamming recognizer is designed and implemented. This recognizer analyzes the received signals to obtain a posterior belief state probability estimate of the current jamming status on each frequency channel. Based on this belief state and a specifically designed reward function targeting anti-jamming effectiveness, the Thompson sampling (TS) algorithm is employed to efficiently solve the MAB model. The Thompson sampling algorithm, through its probability matching mechanism, naturally balances exploration (to discover potentially unjammed channels) and exploitation (to maximize immediate anti-jamming gains using currently known good frequencies). This enables the rapid, adaptive decision-making of optimal or near-optimal frequency transmission strategies in dynamic and uncertain jamming environments.ResultsThompson sampling (TS) demonstrates outstanding performance across all four jamming scenarios: it achieves the highest jamming avoidance success rate with minimal fluctuation and maintains the lowest online selection regret, indicating effective balancing of exploration and exploitation early on leading to stable convergence to the global optimum. Regarding convergence speed and efficiency, TS significantly outperforms ε-greedy, UCB, and DQN, requiring no GPU training (90-100s CPU) and avoiding DQN's heavy reliance on extensive data, long training times (180s GPU), and initial overfitting issues under spot jamming. In action selection, TS dynamically avoids predicted jammed frequencies by learning the jamming strategy in real-time while equally exploring other available channels. This maximizes jamming avoidance while successfully preserving the integrity and randomness of broadband transmission, thereby ensuring the radar's capability to measure target’s distance and velocity.ConclusionsThe TS algorithm demonstrates superior performance across all key metrics: countermeasure effectiveness characterized by high success rate and low regret, learning efficiency evidenced through fast convergence and low computational overhead, and strategic efficacy enabling intelligent jamming avoidance while maintaining detection capability.}
}