Graph-Enhanced Multi-Agent Collaborative Reconfiguration Decision-Making Method for Maritime Unmanned Swarms

ObjectiveTo address the requirement of restoring system functionality after communication link interruptions or node damages in maritime unmanned swarms, this paper investigates a reconfiguration decision-making method that minimizes post-disturbance cumulative performance loss under dynamic topological conditions. MethodsEach unmanned platform in the swarm is treated as an independent decision-making agent, and the reconfiguration problem is formulated as a decentralized partially observable Markov decision process (Dec‑POMDP), under which each platform makes decisions independently based on its local observations. A graph‑enhanced Q‑value decomposition method, termed Graph‑QMIX, is proposed, which employs a graph convolutional network (GCN) to encode platform state features and dynamic communication adjacency into a graph‑level embedding and incorporates it into the QMIX mixing network. This enables the joint value factorization to dynamically adjust the aggregation weights of individual utilities according to the current perturbed topology, thereby achieving collaborative reconfiguration decisions with real‑time perception of the topological structure. ResultsIn simulation environments with swarm sizes ranging from 20 to 100 nodes, Graph‑QMIX improves the comprehensive reconfiguration effectiveness by 41.7% over standard QMIX, and maintains stable performance (comprehensive resilience effectiveness≥0.90) under various disturbance types and intensities. ConclusionsThe experimental results demonstrate that graph‑structured encoding effectively unleashes the inherent resilience potential of the system, significantly enhances the coordination efficiency of multi‑agent scheduling of residual capacities and standby links, and provides a structurally aware and feasible decision‑making approach for collaborative reconfiguration of maritime unmanned swarms.