Dynamic Target Attack for UAV Swarm: A MARL-Based Maneuver Decision-Making Approach

This is the authors' abstract. We don't add key points for this paper.

In the mission execution of unmanned aerial vehicle (UAV) swarms, target allocation and maneuvering decision-making are typically studied separately. This paper integrates these two critical aspects by establishing a coalition formation model for UAV swarms during target selection, a survivability model for dynamic aerial targets under UAVs attack, and a smooth potential field-based dense reward function to coordinate UAV pursuit of targets and collision avoidance. The proposed framework enables UAV swarms to achieve optimal target-oriented maneuvering decisions. Furthermore, a multi-UAV decision-making architecture is designed based on a centralized training with decentralized execution (CTDE) framework, leveraging an improved multi-agent reinforcement learning(MARL) algorithm to ensure efficient and coordinated attacks on multiple targets. The experimental results demonstrate that our approach significantly improves the operational efficiency and adaptability of UAV swarms in dynamic environments.