Autonomous Maneuver Decision for Unmanned Fighters in Multi-Agent BVR Air Combat

In this paper, the issue of autonomous maneuver decision-making in multi-agent unmanned fighter air combat including missile evasion is considered. Firstly, the model of multi-agent unmanned fighter air combat is established by reducing dimensions. The reward function of the MATD3 algorithm model suitable for air combat is designed to realize multi-agent air combat maneuver decisions. Considering the instability in the training process of the MATD3 algorithm in addressing multi-agent continuous action space problems, especially in BVR air combat, an improved multi-agent selective experience replay method is proposed to enhance training stability. Finally, the multi-agent air combat battles are simulated on DCS World and the effectiveness of the MBMATD3 algorithm with the improved experience replay method is demonstrated.