A runway exit prediction model with visually explainable machine decisions

Abstract

A growing number of machine learning (ML) enabled tools and prototypes have been developed to assist air traffic controllers (ATCOs) in their decision-making process. These ML tools can facilitate faster and more consistent decisions for traffic monitoring and management. However, many of these tools utilize models, where machine made decisions are not readily compre- hensible to ATCO. Hence, it is pertinent to develop explainable ML model-based tools for ATCO to manage the inherent risks of using ML model-based decisions. This research investigates visually- explainable ML models for runway exit prediction for better runway management. Specifically, this research adopts local interpretable model-agnostic explanations (LIME) on XGBoost, where machine- made decisions for runway exit prediction are visualized. XGBoost achieved a classification accuracy of 94.35%, 94.17% and 80.87% on the three types of aircraft studied here, respectively. When the LIME parameters are analyzed, Lime shows the contribution of the features for each aircraft corresponding to a particular runway exit. Furthermore, the visual analysis can inform decision makers about the sources of uncertainty in runway exit prediction. Thus, this work paves the way to explainable ML-based prediction of runway exits, where the visually explainable machine decisions can provide insights to ATCO for effective runway management and planning of arrivals and departures. An interactive interface which visualizes machine decisions for runway exit prediction is also developed as a prototype in this paper.