Runway Configuration Assistance: Offline Reinforcement Learning Method for Air Traffic Management

Abstract

Runway configuration management deals with the optimal selection of runways and their direction of operation for aircraft arrivals and departures. The configurations are chosen based on the traffic, surface winds, and other meteorological conditions that are complex to model and difficult to predict. In this paper, we develop the runway configuration assistance tool, an automated approach based on offline model-free reinforcement learning (RL) that provides decision support for air traffic controllers (ATCOs). The proposed tool processes historical data of interest, including decisions made regarding the runway configuration, and their subsequent outcome, to identify a policy that encourages good decisions. The policy search is guided by an appropriately chosen weighted multi-objective utility function (e.g., based on maximizing traffic throughput, minimizing transit times on the surface of the airport, and mitigating safety issues such as go-arounds). The proposed tool is validated using data from two major U.S. airports based on performance metrics developed in collaboration with subject matter experts and is compared against several baseline approaches such as the most frequent configuration chosen by ATCOs, supervised learning, and other RL-based approaches.