Deep Reinforcement Learning for Irregular Flight Recovery With Chain Connectivity Constraint

The irregular flight recovery problem (IFRP) arising from emergency-related airport closures presents substantial challenges due to the need to maintain flight chain continuity and make time-critical operational decisions. Conventional approaches, including mixed-integer programming and heuristic methods, are limited by computational complexity and insufficient flexibility when addressing large-scale disruptions. This study introduces a deep reinforcement learning framework that incorporates a heuristic-guided dueling deep Q-network to solve the IFRP with improved efficiency. The main contributions of this work are as follows. Heuristic action pruning uses domain knowledge to prioritize feasible actions such as cancellations and schedule adjustments while ensuring flight chain continuity. An improved dueling deep Q-network architecture decouples global state value from local action advantages to enhance value estimation in high-dimensional state spaces. Heuristic prioritized experience replay accelerates learning by prioritizing experience samples with greater impact on model performance.