Workload Balancing for Airline Flight Dispatchers

Abstract

We study the flight dispatcher scheduling problem with the objective of balancing workload. Flight dispatchers are responsible for flight planning and monitoring, and their duties are characterized by multi-tasking and uncertainty. We propose novel models, solution methods and a detailed analysis.

First, we introduce and analyze the flight dispatcher workload balancing problem. We model dispatcher workload under two objectives: minimizing peak workload and minimizing absolute workload deviation. To manage large problem sizes, we decompose the problem using Lagrangian relaxation into tractable subproblems and derive a closed-form expression for the Lagrangian lower bound of the peak workload objective. We then propose a focus-search-improve genetic algorithm, which integrates genetic search with refinement heuristics. Numerical tests on real-world flight data show that the heuristic finds optimal or near-optimal solutions and outperforms commercial solvers. It is found that minimizing peak workload reduces workload spikes while minimizing absolute deviation distributes workload uniformly.

Second, we incorporate practical and operational constraints such as route-based flight grouping and assignment restrictions based on flight characteristics. We provide a multi-objective model and propose three solution approaches, the weighted-sum method, the epsilon-constraint method, and the non-dominated sorting genetic algorithm II, to capture the trade-off between workload balance and operational efficiency. Computational experiments show that non-dominated sorting genetic algorithm II generates clear Pareto fronts and provides solutions that are comparable to those found by exact methods.

Third, we focus on the absolute deviation workload balancing problem which achieves balanced schedules but is computationally challenging. We propose a neural network-assisted column generation algorithm that predicts negative reduced cost solutions of the Lagrangian subproblems. Preliminary results demonstrate the viability of the approach in speeding up the column generation algorithm.