DyVertexGAT: A Novel GNN-based Target Assignment Policy for Multi-Aircraft Combat Decision

In high-fidelity close-range air game, the stochastic nature of maneuvering engagements renders traditional optimization methods incapable of real-time solution search for dynamic target allocation. To address this, we decompose the multi-aircraft engagement into a hierarchical decision framework consisting of a target allocation layer and a maneuver execution layer. We first formulate a multi-stage dynamic target assignment problem using mixed-integer programming and model it as a Markov game. To handle the time-varying number of targets, we represent the engagement as a temporal dynamic graph and propose the Dynamic Vertex Graph Attention Network (DyVertexGAT), which uses graph attention mechanisms to compute node importance and generate allocation policies that naturally adapt to variable graph sizes. We further introduce Self-Play Proximal Policy Optimization (SPPPO) to iteratively refine the policy through self-play. Simulation results show that DyVertexGAT outperforms a fixed-scale graph attention baseline trained with PPO, achieving a 54% win rate compared to 40% for the baseline. Behavioral analysis confirms that the learned policy makes efficient use of ammunition and exhibits tactically sound decision-making. This work validates the effectiveness of DyVertexGAT and SPPPO for dynamic target allocation in uncertain, adversarial environments.