Hierarchical Multi-Armed Bandit-Based Decision Optimization for Robust Radar Detection Against Agile Jammer

Robust and accurate target detection capability in complex confrontation environments is of great significance for radar intelligence reconnaissance. With the development of agile jamming technologies and equipment, conventional and exist cognitive anti-jamming methods may suffer from jamming suppression performance loss due to strategy mismatch and response lag in such non-stationary jamming environments. This paper jointly optimizes the selection of anti-jamming algorithms and their parameters to achieve more flexible strategy adjustment, thereby solving the strategy mismatch problem. To address the strategies explosion caused by joint-space, we propose a hierarchical multi-armed bandit (MAB) decision method. It first selects the most promising algorithm, then fine-tunes its parameters. Meanwhile information gain ratio is computed from historical interaction data to rank cross-module algorithms and their parameters, enabling hierarchical decomposition of the decision space. The upper confidence bound (UCB) approach is applied on the ranked candidates to balance exploration and exploitation under unknown reward distributions. Moreover, to enhance the algorithm's response to agile jammers, an online anti-jamming decision framework with a memory mechanism is established, which uses coarse jamming change-point information to retrieve interaction histories and accelerate adaptation to new jamming states. Simulation results under agile jamming scenarios show that the proposed method significantly improves both target detection probability and adaptability in non-stationary environments against agile jammers.