Large-Scale Airline Recovery using Mixed-Integer Optimization and Supervised Learning

Abstract

Airlines plan their aircraft and crew schedules using operations research methods. However, these schedules are often disrupted due to the irregular nature of flight operations. Airline recovery is the process in which airlines take various actions to adjust and repair their aircraft routes, crew schedules, and passenger itineraries. This process has a sequential structure in practice, where aircraft recovery is followed by crew and then passenger recovery. Although recovery problems are smaller in scope than their planning counterparts, limited solution timeframes prevent airlines from using a full-scale optimization approach.

This thesis proposes fast solution methods that combine mixed-integer optimization and supervised machine learning techniques to find better solutions to large-scale airline recovery problems than those found with other exact and heuristic approaches. Our approach reduces the solution space for a given disruption by adding constraints (cuts) based on the patterns discovered in the solutions to historical disruptions. The model with the added cuts is solved using mixed-integer optimization solvers.

During the day of flight operations, the available time for airlines to handle disruptions may vary. The overall solution process we propose allows parameter tuning to match the extent of solution space reduction with the available solution time. This feature helps the proposed methods to effectively navigate the trade-off between solution quality and runtime. Our computational studies are conducted using real flight and crew schedules from major US airlines with more than 2,500 daily flights. Experiments demonstrate that our approach can generate solutions of significantly higher quality than benchmark methods.

We use tree-based classification methods to predict recovery decisions. Due to their interpretable structures, we are able to discover insights into the attributes of effective recovery decisions. We demonstrate that these insights can be incorporated into currently used heuristic-based airline recovery processes to improve solution quality by up to 15%.