To address the perimeter defense problem for multi-agent systems, this paper proposes a decoupled strategy based on local observation matching to overcome the curse of dimensionality and convergence difficulties in traditional multi-agent reinforcement learning (MARL). This strategy decomposes the perimeter defense algorithm into two components. First, at the task assignment layer, a distributed Hungarian matching mechanism is introduced to achieve communication-free implicit coordination. Second, at the maneuver control layer, a Multi-Agent Soft Actor–Critic (MASAC) framework incorporating parameter sharing is constructed. Simulation results demonstrate that the proposed strategy outperforms baselines in both convergence speed and training rewards. Furthermore, it exhibits zero-shot generalization capabilities, enabling direct extension to large-scale perimeter defense missions.
Huang et al. (Fri,) studied this question.