针对雷达系统面临的干扰场景复杂多变、人工设计抗干扰策略性能难以保证以及实时性不高的问题,构建了基于深度强化学习的智能决策生成模型,设计了有针对性的动作集、状态集和奖励函数。同时提出了基于双深度Q 网络(DDQN)的决策网络训练算法,用于克服深度Q 网络(DQN)算法中目标网络与评估网络相耦合导致Q 值的过估计。仿真结果表明:与DQN、Q 学习、人工制定策略与遍历策略库等方法相比,文中所设计的智能决策模型和训练方法对干扰的抑制效果好,泛化能力更强,反应时间更快,有效地提升了雷达自主决策能力。

In order to solve the problems faced by radar system such as complex jamming scenes, low reliability and bad real-time performance, an intelligent decision generation model is constructed based on Deep Reinforcement Learning, where targeted action set, state set and reward function are designed. After that, a decision network training algorithm based on double deep Q-network is proposed to overcome the problem of Q value over estimation which caused by the coupling of target network and evaluation network in Deep Q-network (DQN). The simulation results show that, compared with DQN, Q learning and traversal algorithm, the intelligent decision model and training method designed in this paper have better interference suppression effect, stronger generalization ability and faster response time, and effectively improve the radar independent decision-making ability.

TN972

国家自然科学基金资助项目(62171220)