RL-AirSched: A Feasibility-Aware Reinforcement Learning Framework for Dynamic Airport Slot Rescheduling

Airport slot rescheduling is a complex sequential decision-making problem in which operational disruptions, limited airport capacity, and stringent feasibility constraints require timely rescheduling decisions. While conventional optimization techniques have achieved strong performance in strategic slot allocation, their reliance on repeated re-optimization limits their adaptability to continuously changing operational conditions. Although reinforcement learning (RL) has emerged as a promising paradigm for dynamic scheduling, its application to airport slot rescheduling remains hindered by the large number of infeasible actions that dominate the decision space and impede efficient policy learning. This paper proposes RL-AirSched, a feasibility-aware reinforcement learning framework for dynamic airport slot rescheduling. RL-AirSched integrates constrained candidate-action generation, binary action masking, and a multi-objective reward formulation to guide policy learning within a feasibility-aware scheduling environment. By restricting exploration to a feasibility-filtered set of flight-slot assignments, the framework enables the learning agent to balance conflict mitigation, displacement minimization, fairness preservation, and successful flight reassignment within a unified sequential decision-making process. The proposed framework is evaluated using three airport scheduling datasets representing progressively increasing operational complexity and is compared with Random Agent and Greedy Heuristic benchmark strategies. Comprehensive experiments include comparative performance evaluation, scalability analysis, multi-seed robustness assessment, and one-way ANOVA statistical validation. The results demonstrate that RL-AirSched consistently produces effective scheduling policies within the evaluated feasibility-aware environments, achieves all 30 successful flight reassignments in the medium-scale scheduling environment with zero variance across ten independent training runs, maintains competitive performance in the larger tested scheduling environments, and demonstrates scalability across the three evaluated synthetic problem sizes. The findings demonstrate that feasibility-aware action-space engineering provides an effective mechanism for improving reinforcement learning in constrained scheduling problems. Beyond airport slot rescheduling, the findings provide experimental evidence supporting feasibility-aware reinforcement learning for constrained sequential decision-making, while validation on unseen operational configurations and real airport data remains necessary to establish broader generalizability and practical applicability.