Combat Target Classification Method Based on Human-Machine Agent with Double-Q Policy Under Weak Labeling Conditions

The task of combat target classification, as an important component of command automation systems, is not only a prerequisite for battlefield situational awareness and threat assessment, but also an important basis for battlefield decision-making. However, the special nature of military confrontation dictates that sample data in this field is small, and those with high-quality labels are even scarcer, while their label annotation work is exceptionally expensive. Therefore, based on reinforcement learning DQN (Deep Q-Network) and Q-learning, we propose a double-Q policy update method (DQQN) for the task of target classification under weak annotation in combat environments to improve the sampling efficiency of target data. At the same time, combined with active learning, a human-machine agent MTCARq_H-M is proposed, which can intelligently infer whether to automatically perform target classification or introduce manual annotation. Then we carried out application analysis of the proposed model based on the replay dataset of a public wargaming platform, and the results demonstrate that the proposed model MTCARq_H-M not only offers significant labour cost savings compared to the purely supervised model Supervised, the traditional active learning algorithms QBC (Query By Committee) and Uncertainty, but also achieves more competitive combat target classification accuracy.