Stochastic models of Ground Delay Program implementation for prediction, simulation, and insight

Abstract

We use historical data to build two types of stochastic model of hourly Ground Delay Program (GDP) implementation for the three main airports near New York City: Newark Liberty International, LaGuardia, and John F. Kennedy International. The models predict the probability that a GDP will be initial ized or canceled in a given hour based on hundreds of features describing the situation at the airport, including features describing forecasted weather conditions and scheduled traffic. One is a regularized logistic regression model that ignores system dynamics and the other model is based on inverse reinforcement learning, so it considers system dynamics and the impact that GDP implementation actions have over time on some metrics. We evaluate the models based on two objectives: their ability to predict and simulate GDP implementation decisions in historical test data sets. As is expected based on the motivation for and objective of each type of model, regularized logistic regression models make superior predictions while simulations controlled by inverse reinforcement learning models produce average metric values that more closely match those produced by historical GDP implementation decisions. Finally, we draw insights about GDP implementation from the trained model parameters. For example, parameters of both models suggest while weather conditions and values of performance metrics such as airborne delay play a role in GDP decision making, parameters related to the predictability or continuity of existing GDP plans are also important.