Learning Scalable Task Assignment with Imperative-Priori Conflict Resolution in Multi-UAV Adversarial Swarm Defense Problem

The multi-UAV adversary swarm defense (MUASD) problem is to defend a static base against an adversary UAV swarm by a defensive UAV swarm. Decomposing the problem into task assignment and low-level interception strategies is a widely used approach. Learning-based approaches for task assignment are a promising direction. Existing studies on learning-based methods generally assume decentralized decision-making architecture, which is not beneficial for conflict resolution. In contrast, centralized decision-making architecture is beneficial for conflict resolution while it is often detrimental to scalability. To achieve scalability and conflict resolution simultaneously, inspired by a self-attention-based task assignment method for sensor target coverage problem, a scalable centralized assignment method based on self-attention mechanism together with a defender-attacker pairwise observation preprocessing (DAP-SelfAtt) is proposed. Then, an imperative-priori conflict resolution (IPCR) mechanism is proposed to achieve conflict-free assignment. Further, the IPCR mechanism is parallelized to enable efficient training. To validate the algorithm, a variant of proximal policy optimization algorithm (PPO) is employed for training in scenarios of various scales. The experimental results show that the proposed algorithm not only achieves conflict-free task assignment but also maintains scalability, and significantly improve the success rate of defense.