Publication
Study summary
This study develops an application-driven model-free reinforcement-learning framework for tactical demand-capacity balancing at São Paulo/Guarulhos International Airport. The policy jointly manages runway configuration, arrival-departure capacity allocation, and miles-in-trail separation using Maskable PPO, incremental separation actions, action smoothing, and curriculum training that combines historical operations with forecast-based stochastic scenarios. The evaluation shows that forecast-based training is important for learning a non-trivial separation policy. Relative to capacity management without separation control, the approach reduces average queue-related cost while smoothing severe congestion peaks, but introduces additional spacing delay. The paper therefore treats miles-in-trail control as an operational trade-off rather than universally lower generalized cost, and finds that targeted separation at or below 10 NM is generally sufficient on the most severe days under the SBGR case assumptions.
Keywords: traffic flow management; demand-capacity balancing; machine learning; reinforcement learning; miles-in-trail; optimization.
BibTeX
Please cite the published journal article through its DOI.