Imitation Learning of Jamming Strategies Under Low-Quality Expert Database Conditions

Abstract

This paper proposes an improved generative adversarial imitation learning (GAIL) method for multi-functional radar (MFR) jamming decision-making, addressing the poor initial jamming effect of traditional deep reinforcement learning (DRL) and the high database quality requirement of basic GAIL. By introducing the behavior model (BM) and negative database (ND), the proposed method can improve the agent's imitation accuracy of the expert policy in low-quality data environments, avoiding learning ineffective jamming strategies. The agent strives to mimic the expert strategy during offline training through continuous updates. In online application, the trained Actor network directly informs jamming pattern decisions. The proposed method is validated through comparison with four other algorithms, including random strategy and basic GAIL. It effectively reduces MFR entry into high-threat states and improves aircraft survivability, offering a robust solution for enhancing aircraft survivability in complex electromagnetic environments.