Policy-based Bayesian-MILP optimization framework for airport departure sequencing under uncertainty

Efficient departure sequencing is central to reducing delays, fuel burn, and environmental impacts at congested airports. Although Mixed Integer Linear Programming (MILP) can deliver optimal solutions, it is often used as a reactive, one-shot plan with static or manually tuned multi-objective weights, and thus struggles with time-varying traffic and operational uncertainty. We propose a proactive, policy-based framework that treats a MILP as a tunable decision policy rather than a fixed schedule. The novelty is a smart policy search performed before operations. Using sequential Monte Carlo simulation and a grid search over a normalized multi-objective cost (delay, fuel, emissions, makespan, and position stability), the framework identifies the best weight vector ( 𝜃 ∗ ) for a given traffic scenario. To overcome discrete search limits, we introduce an Advanced Multi-Stage Bayesian Optimizer (AMS-BO) that explores, zooms, and refines the parameter space to uncover non-linear trade-offs. Experiments on 50 traffic samples from the Antalya Airport surface layout show that no single weighting scheme is universally optimal; the winning policy varies by scenario and is tailored to each sequence’s dynamics. Without proactive tuning, conventional benchmarks can produce pathological outcomes where improving one metric degrades others. In contrast, the policy-based MILP reduces average delays by 36%, hold-fuel burn by over 80%, and delay-related emissions by 69% under pushback uncertainties, relative to a static schedule. With stable sub-second solves in a rolling-horizon loop, the framework can complement existing Departure Manager (DMAN) and Terminal Flight Data Manager (TFDM) systems, enhancing sustainability and resilience in capacity-constrained environments.