A machine learning approach for solution space reduction in aircraft disruption recovery

Abstract

Aircraft recovery, a critical step in airline operations recovery, aims to minimize the cost of disrupted aircraft schedules. The exact methods for aircraft recovery are computationally expensive and operationally infeasible in practice. Heuristics and hybrid approaches offer faster solutions but have inconsistent solution quality, often leading to large losses. We propose a supervised machine learning approach to accelerate aircraft recovery by pruning the solution space of the optimization problem. It leverages similarities with previously solved problem instances through an offline model-training phase, identifies components of the optimal solutions for new problem instances in the online phase, and links them to the optimization model to rapidly generate high-quality solutions. Computational results, from multiple historical disruption instances for a large US airline, demonstrate that this approach significantly outperforms exact methods on computational runtime while producing similarly high-quality solutions. It also outperforms existing heuristics due to its ability to prune solution spaces in a more principled manner, leading to higher quality solutions in similarly short runtimes. For a runtime budget of two minutes, our approach provides a solution within 1.5% of the true optimal cost, resulting in an average daily saving of over $390,000 compared to all existing approaches. The main drivers of these improvements are explainable in terms of key airline operational metrics and are validated through extensive sensitivity and robustness tests.