Abstract: This study addresses the need for multi-UAV operations in complex and dynamic environments to perform group target expulsion and encirclement tasks, and proposes an intelligent decision-making algorithm based on the Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3). A task scenario close to real combat and a UAV kinematic model are constructed, while the artificial potential field method is introduced to design the maneuvering strategy of offensive UAVs. On this basis, a multi-dimensional reward function integrating guided rewards and sparse rewards is designed, combined with safety distance and boundary constraints to enhance the safety and robustness of UAV collaborative operations. The algorithm adopts a centralized training and decentralized execution framework, and employs dual-Critic networks and a delayed update mechanism to overcome Q-value overestimation and training instability. Simulation results show that the proposed method achieves efficient expulsion and encirclement in both 2-vs-1 and 3-vs-1 scenarios, with a 13.3% improvement in convergence speed compared with the MADDPG algorithm. The findings demonstrate that the proposed approach significantly enhances the autonomous decision-making and cooperative capabilities of UAV swarms, providing a feasible new solution and technical support for future aerospace defense and group-to-group combat.