关键词: 舰载机, 保障作业, 调度优化, 强化学习, 深度学习

Abstract: In aircraft carrier flight deck scheduling operations, model-free reinforcement learning (MFRL) is constrained by the precision of physical environment modeling under dynamic deck conditions. In contrast, model-based reinforcement learning (MBRL) faces challenges of high computational complexity and convergence difficulties due to the interdependent co-optimization between the environment model and the decision-making model during iterative training. To address these issues, this paper proposes a hybrid reinforcement learning framework (MB-MF) that integrates model-based and model-free characteristics. First, a deep neural network-based deck environment model is trained using historical scheduling data to accurately predict state transitions within a minimal tolerance range. Then, the converged environment model is embedded into the interactive environment in place of the real environment, where a scheduling agent is trained using the Deep Q-Network (DQN) algorithm, thereby decoupling environment model learning from policy optimization. Experimental results demonstrate that, compared to MFRL using the physical environment, the proposed method achieves a performance gap of only 4% without requiring precise modeling. Moreover, it reduces the aircraft sortie time by 34% relative to the MBRL baseline. In resource-constrained scenarios, the decision-making speed is nearly 300 times faster than that of heuristic methods, while the scheduling quality is reduced by only 17%.

Key words: carrier-based aircraft,, deck operations,, complex job scheduling,, reinforcement learning,, deep learning