Deep Reinforcement Learning based Autonomous Air-to-Air Combat using Target Trajectory Prediction

This study designed an intelligent control system for autonomous air-to-air combat and verified it in a realtime flight simulation. Previous studies of aerial combat have required significant effort to design agile control actions for different engagement conditions. In this work, optimal flight control under random engagement conditions was performed by using reinforcement learning and recurrent neural networks. A target trajectory was predicted using Sequence-to-Sequence model with LSTM, for occupying an advantageous location from an enemy aircraft in a close engagement. In addition, this study proposed an algorithm with improved performance compared to the existing algorithm. The result of the study confirmed that the maneuvers of trained agent were similar to the performance of human pilots and the future position of the enemy was tracked by own ship aircraft.