Surrogate Modeling of Agent-Based Airport Terminal Operations

Abstract

The airport terminals are complex sociotechnical systems, which are difficult to understand and their behavior is hard to predict. Hence, an agent-based model, the Agent-based Airport Terminal Operation Model (AATOM), has been designed to represent and analyze diverse airport terminal processes, actors, their behavior and interactions. The main issue with such models is the large computational requirements for simulating detailed processes, making it computationally inefficient. Furthermore, the dynamics of such models are difficult to understand. Therefore, the goal of this research is to approximate the dynamics of AATOM by a surrogate model, while preserving the important system properties. A methodology is suggested for training and validating a surrogate model, based on the Random Forest algorithm. The trained surrogate model is capable of approximating the AATOM simulation and identifying relative importance of the model variables with respect to the model outputs. Firstly, the results obtained contain an evaluation of the surrogate model accuracy performance, indicating that the surrogate model can achieve an average accuracy of 93% in comparison to the original agent-based simulation model. Nonetheless, one indicator, the number of missed flights, has shown to be more difficult to predict, with an average accuracy of 83%. Secondly, the results show that the airport resource allocation has an important impact on the efficiency of the airport terminal, with the two most important variables being the number of desks at the check-in and the number of lanes at the checkpoint. Last, the developed surrogate model was compared with a second Artificial Neural Network-based surrogate model built for the same agent-based model.