Abstract

International audience; In this paper, we assess the performance of a workload model trained on a subset of sectors, focusing on how it generalizes on fresh sectors. The model of the air traffic controller workload is learned from historical data made of workload mesurements extracted from past sector operations and ATC complexity measurements computed from radar records and airspace data (sector geometry). The workload is assumed to be low when a given sector is collapsed with other sectors into a larger sector, normal when it is operated as is, and high when it is split into smaller sectors assigned to several working positions. This learning problem is modeled as a classification problem where the target variable is a workload category (low, normal, high) and the explanatory variables are the air traffic control (ATC) complexity metrics. In previous work, we compared several classifiers on this problem. The models were trained on one week of traffic, and their generalization performance was assessed on another week of traffic, using the same sectors in both the training and test sets. In the current work, we examine if models learned on a specific set of sectors can be performant on any other sector, or not. We also give a closer look at how the workload varies with the ATC complexity measures in our data, using bagplots of the data points for a few sector instances. The results allow us to better understand the strengths and limits of our data-driven model.

Abstract

International audience; In this paper, we compare several machine learning methods on the problem of learning a model of the air traffic controller workload from historical data. This data is a collection of workload mesurements extracted from past sector operations and of ATC complexity measurements computed from radar records and airspace data (sector geometry). We assume that the workload is low when a given sector is collapsed with other sectors into a larger sector, normal when it is operated as is, and high when it is split into smaller sectors assigned to several working positions. This learning problem is modeled as a classification problem where the target variable is a workload category (low, normal, high) and the explanatory variables are the air traffic control (ATC) complexity metrics. Several classifiers are compared on this problem: linear dis-criminant analysis, quadratic discriminant analysis, naive Bayes classifiers, neural networks, and gradient boosted trees. The performance of these models is assessed on a separate test set. The best methods show a rate of correct predictions around 82%.

Abstract

International audience; In this paper, we introduce an Ant Colony System algorithm which finds optimal or near-optimal sequences of airspace partitions, taking into account some constraints on the transitions between two successive airspace configurations. The transitions should be simple enough to allow air traffic controllers to maintain their situation awareness during the airspace configuration changes. For the same reason, once a sector is opened it should remain so for a minimum duration. The Ant Colony System (ACS) finds a sequence of airspace configurations minimizing a cost related to the workload and the usage of manpower resources, while satisfying the transition constraints. This approach shows good results in a limited time when compared with a previously proposed $A$ * algorithm on some instances from the french air traffic control center of Aix (East qualification zone) where the $A$ * algorithm exhibited high computation times.00