The revenue and performance of a flight schedule is highly dependent on its robustness. Robustness can be thought of as the flexibility available within a schedule to enable a user to recover from disruptions to that schedule. A more detailed description of robustness can be found in [8]. A lack of robustness in a schedule might result in one delay leading to a series of knock on delays which can seriously affect the smooth operation of an airline. In order to mitigate against this, airline companies want to have accurate estimates of a schedule’s robustness before that schedule is brought into operation. Regular evaluation during the scheduling process enables schedule operators to improve the schedule quality by adding slack where necessary and by incorporating strategic patterns – for instance a favorable slack distribution into the schedule. A more detailed description of robust scheduling within the airline industry can be found in [1, 2]. Our full paper will present custom-built and generic models which have been developed to evaluate schedule robustness. This work builds upon some of the conclusions arising from research carried out by Lee et al. [7], Schumacher [10], Rosenberger et al. [9] and Farrington et al. [5] on stochastic modeling which provides schedule operators with a detailed what-if analysis of the schedule. However, although stochastic models provide good estimates of schedule robustness, operators often find them unsatisfactory because of their heavy computational load. This prevents them from performing schedule evaluations on a regular basis. Our method represents a much quicker approach to help operators to build better models. Moreover, our initial experiments suggest that the approach is also more accurate but this hypothesis will be rigorously evaluated and discussed in the full paper. We explore a prediction of robustness which is based upon certain robustness features which are drawn from schedule operators at KLM and from recent papers in the literature [1, 2]. This feature-based prediction enables us to forecast schedule robustness. The robustness features in the flight schedule which we are exploring include: • slack distribution • likeliness of delay propagation • number of redundant paths [1] • number of swap options • presence of strategic patterns • number of meeting points in the schedule The goal is that these features (and others) will be identified and translated into mathematical parameters called explanatory variables. The selection of a relevant subset is carried out using wrapper feature selection [6] on historic real world data from KLM. In this iterative process an induction algorithm [6] is used to train a classifier using a selected subset of features as input. Model evaluation is carried out using n-fold cross-validation [6] on the training set. The iterative process finishes once the search
[1]
Ron Kohavi,et al.
Wrappers for Feature Subset Selection
,
1997,
Artif. Intell..
[2]
Bruce Schumacher.
Proactive flight schedule evaluation at Delta Air Lines
,
1999,
WSC '99.
[3]
Thomas G. Dietterich.
Machine-Learning Research Four Current Directions
,
1997
.
[4]
George L. Nemhauser,et al.
A Stochastic Model of Airline Operations
,
2002,
Transp. Sci..
[5]
J. D. Landa Silva,et al.
MAKING AIRLINE SCHEDULES MORE ROBUST
,
2003
.
[6]
Gabriel A. Wainer,et al.
Proceedings of the 2016 Winter Simulation Conference
,
2016
.
[7]
Leo Breiman,et al.
Bagging Predictors
,
1996,
Machine Learning.
[8]
John-Paul Clarke,et al.
Approaches to incorporating robustness into airline scheduling
,
2000
.
[9]
Roel Leus,et al.
The generation of stable project plans
,
2004,
4OR.