Using Machine Learning to Include Planners’ Preferences in Railway Crew Scheduling Optimization

In crew scheduling, optimization models can become complex when a large number of penalty terms is included in the objective function to take planners’ preferences into account. Planners’ preferences often include nonmonetary aspects for which both the mathematical formulation and the assignment of appropriate penalty costs can be difficult. We address this problem by using machine learning to learn and predict planners’ preferences. We train a random forest classifier on planner feedback regarding duties from their daily work in railway crew scheduling. Our data set contains over 16,000 duties that planners labeled as good or bad. The trained model predicts the probability that a duty is perceived as bad by the planners. We present a novel approach to replace the large construct of penalty terms in a crew scheduling optimization model by a single term that penalizes duties proportionally to the predicted probability of being assessed as unfavorable by a planner. By integrating this probability into the optimization model, we generate schedules that include more duties with preferred characteristics. We increase the mean planner acceptance probability by more than 12% while only facing a marginal increase in costs compared with the original approach that utilizes a set of multiple penalty terms. Our approach combines machine learning to detect complex patterns regarding favorable duty characteristics and optimization to create feasible and cost-efficient crew schedules. Supplemental Material: The online appendix is available at https://doi.org/10.1287/trsc.2022.1196 .