Integration of Reinforcement Learning into Planning strategies for Flight Diversion Support

This Thesis explores the application of Reinforcement Learning (RL) to optimize flight path computations for Airbus A320 airliners, focusing on enhancing efficiency and adaptability under uncertainty. Current methods rely on Combinatorial Optimization and computationally intensive simulations to estimate optimal routes. By training an RL Agent to precompute ideal paths based on atmospheric data, the approach reduces the search graph size, significantly speeding up route optimization. Although optimality is not guaranteed, empirical results show that fuel consumption is almost always on a par with the unconstrained solver or slightly off ($\leq$ 1\%), while at the same time the computation speed can be reduced by up to 50\% as compared to standard combinatorial optimization techniques. Additionally, the intrinsic resilience to noisy and imprecise data of Machine Learning addresses at best the challenges posed by unreliable weather predictions, thus producing more robust solutions than traditional methods. The Thesis discusses with great detail the theoretical framework, the different implementation strategies, the adopted testing procedures and the obtained results, also reporting on discarded approaches as well as future perspectives.