Deep reinforcement learning for carrier-based aircraft flight deck operations scheduling problem

Flight deck operations scheduling is an NP-hard combinatorial optimization problem, where traditional methods face a critical trade-off between computational efficiency and solution quality. To address this challenge, we propose a deep reinforcement learning framework integrated with graph neural networks to optimize this process. The problem is formulated as a Markov decision process, allowing the scheduling agent to generate schedules directly from the environment state. Our analysis identifies that a softmax exploration strategy combined with a discount factor of 1.0 provides a robust configuration for general applicability. Experimental results demonstrate that the agent outperforms traditional priority dispatching rules regarding solution quality. Compared to meta-heuristic algorithms, our well-learned agent achieves competitive performance on small-scale problems and demonstrates superior search capabilities on large-scale instances. Notably, the agent reduces decision-making time from dozens of minutes required by meta-heuristics to mere seconds, while producing high-quality solutions that meet real-time operational demands.