关键词: 航班着陆调度, 组合优化, 强化学习, 策略梯度, 注意力机制

Abstract: The Aircraft Landing Problem (ALP) is a high-dimensional, non-linear combinatorial optimization challenge characterized by stringent safety separation constraints. To address the "curse of dimensionality" faced by swarm intelligence algorithms in large-scale traffic scenarios—which struggle to meet real-time requirements—and the convergence difficulties of traditional Reinforcement Learning (RL) when handling hard constraints, this paper proposes a Guided Policy Gradient (GPG) method. First, a sequence decision network based on the Attention Mechanism is constructed. Its multi-head attention layers are utilized to capture complex spatio-temporal correlations between flights, enabling precise perception of the global scheduling state. Second, to tackle high dimensionality, a Dynamic Restricted Candidate Mask (DRCM) strategy is introduced. By applying target-time criteria to filter the most relevant flight subsets in real-time, the high-dimensional action space is effectively compressed into a fixed-size rolling decision window, significantly reducing policy search costs. Finally, a Relaxed Guidance mechanism is proposed to transform discrete hard separation constraints into smooth gradient guidance paths, effectively resolving the non-differentiable issues within the ALP decision space. This is combined with an Exponential Moving Average (EMA) baseline mechanism to reduce policy gradient variance and enhance training stability under massive stochastic samples. The proposed method is validated using the OR-Library dataset. Experimental results demonstrate that the model outperforms the First-Come-First-Served (FCFS) strategy, the Displacement and Location Planning (DALP) Algorithm, Scatter Search (SS), and the Bio-inspired Algorithm (BA). Specifically, the solution accuracy is improved by an average of 44.58%, 20.69%, 3.16%, and 0.84%, respectively. On small-scale datasets, the performance is comparable to the Hybrid Particle Swarm Optimization with Local Search (RH-HPSO-LS). On large-scale datasets, the proposed algorithm maintains high precision while demonstrating superior time efficiency compared to existing methods, indicating its ability to balance computational efficiency with constraint robustness.

Key words: aircraft landing sheduling, combinatorial optimization, reinforcement learning, policy gradient, attention mechanism