Abstract: To analyze and predict the opponent’s intentions and future targets scientifically in real-time adversarial tasks, a multi-agent recognition algorithm was proposed based on data driving. Firstly, taking a feature extraction method with automata, the arithmetic was arranged to obtain position and task information from samples for planning. Then, transforming the planning recognition into a multi-classification problem, a multi-classification model was developed based on extreme gradient boosting (XGBoost) to start with a single agent. And then, considering collaboration potential among agents, the density-based spatial clustering of applications with noise (DBSCAN) algorithm was employed for multi-agent clustering and collaborating. And a multi-classification model was then developed for clustered multi-agent systems to predict goals within the same cluster. After identifying the opponent's targets, an encirclement and stop strategy was designed based on game theory. Finally, a non-cooperative dynamic game model was constructed, and the Nash equilibrium solution was used to derive the optimal strategy. Simulation results demonstrate the effectiveness of the proposed algorithm.