My IVADO scholarship provided me with financial support that enabled me to devote myself to my master’s degree full-time.
In the field of civil aviation, flight planning is a real brain-teaser. Not only do airlines have to take into account passengers and their baggage, but also the crew of each aircraft. Unlike the former, who are simply going from Point A to Point B, the three or four flight attendants remain assigned to an aircraft until it returns to its home airport. These flight sequences last an average of four days. They are called rotations.
An employee’s monthly schedule consists of a succession of rotations interspersed with days of time off or vacation. Assignment of these rotations depends on many criteria, such as the languages spoken by the employee, his or her preferences and visas. Generating all these schedules requires a lot of calculations and therefore takes a great deal of time. This is why airlines, whose flight crews are their second-largest expenditure after fuel, put a lot of effort into optimizing these calculations.
Alice Wu has been working on this optimization as part of her master’s research in applied mathematics. She studied several thousand past rotations at one airline, and how the algorithm used by the company assigned these rotations to its employees to build individual and personalized schedules. Her goal was to accelerate the process by making it easier for the algorithm to work.
There are a lot of parameters to take into account: constraints, collective agreements, number of hours worked, number of hours in the air.
She sought to make the algorithm more intuitive in order to reduce the number of rotations taken into account at the beginning of the calculation. Instead of a pool of 5,000 potential rotations, the algorithm had to consider only the 500 most likely. To create this intuition, she used machine learning. She then checked whether the schedule that had actually been assigned to the employee in the past was among the 500 chosen by the algorithm.
Finally, she evaluated the efficiency of her model by comparing it with a random selection system. “With randomness, only 10% of the rotations that were actually assigned in the past are found in the top 500. With my model, I managed to reach 40%, which is quite positive.”
The young student’s research has not yet yielded solid results that contain enough simulations on different data and different airlines. Her model is therefore not yet developed enough to be used. After completing her master’s degree in December 2019, she is now continuing her research using more data to further improve her model.