Multi-Stage Dynamic Weapon-Target Assignment Based on Improved Reinforcement Learning

The dynamic weapon-target assignment (DWTA) problem in modern combat environments, characterized by multiple waves and high dynamics, is tackled by incorporating practical scheduling constraints such as weapon availability time, cooldown intervals, and target engagement time windows. To address this problem, we construct a multi-objective optimization model to simultaneously consider threat elimination, base defense, and inter-phase resource coordination, aiming to improve the foresight and overall performance of the scheduling strategy. In terms of algorithm design, we propose an improved Actor-Critic algorithm, where a pointer network is integrated into the Actor module, leveraging its attention mechanism and dynamic masking capability to efficiently select weapon-target pairs under variable-length input conditions. Furthermore, we employ a dual-channel encoding-decoding mechanism to enhance the solution accuracy of the algorithm. The experimental section includes validation of the objective function, performance comparison with baseline algorithms, generalization tests, and ablation analysis of the attention mechanism and regularization components. The results validate the superior optimization capability and adaptability of the proposed model and algorithm in complex dynamic environments.