Graph Neural Network and Deep Reinforcement Learning Driven Kill Chain Construction: Supporting Cross‐Domain Task Allocation Decision System

At present, the decoupling and distributed deployment of combat resources result in low real‐time performance and poor reliability of autonomous task allocation decision‐making systems during resource scheduling. Traditional centralized task allocation methods are unable to address these challenges with adequate flexibility and adaptability. To solve the above problems, this paper proposes a kill chain construction method combined with graph neural networks (GNNs) and deep reinforcement learning so as to support the efficient operation of trans‐domain task allocation decision‐making systems. First, to effectively capture the information interactions, interdependent coupling relationships, combat capabilities, and core objectives of heterogeneous combat resources in air–ground cooperative kill chain construction, an RM kill chain model is designed. The theoretical advantages of the reverse‐sequence logic and multilayer architecture are theoretically analyzed under explicit assumptions and experimentally validated from both the construction logic and hierarchical architecture perspectives. Subsequently, tailored to the inherent characteristics of the proposed RM model, the dual‐stage GACP‐KM algorithm is developed. In the first stage, node feature learning and adaptive partitioning are achieved based on GNNs and multi‐Actor‐Critic architecture. In the second stage, the Kuhn–Munkres (KM) algorithm is employed within each partition to achieve optimal construction of kill chains. In particular, a factorized partitioning action space, a partition‐number‐invariant scoring network, and a two‐level reward driven by the final construction quality are designed to tightly couple the learning‐based partitioning stage with the exact matching stage. Finally, the proposed methods are applied and validated in cross‐domain cooperative combat scenarios. Comprehensive experimental results demonstrate the superiority of the RM model, the computational efficiency and solution quality of the GACP‐KM algorithm, and the strong compatibility between the model and algorithm.