Construction of Aircraft Pilot Model by Inverse Reinforcement Learning

In this study, we tried to construct a pilot model by applying Generative Adversarial Imitation Learning, a practical method for imitating expert behavior, to the horizontal steady flight and landing problems of aircraft. Maneuvering data were collected from expert pilots through flight simulation experiments and used Generative Adversarial Imitation Learning to train a controller based on deep neural networks. In constructing the landing model, it was found to be effective to separate and combine the approach and flare modes. We successfully obtained the pilot model which captures the characteristics of expert pilot. Furthermore, we evaluated the pilot model's performance through Monte Carlo simulations, providing a quantitative assessment of its applicability and reliability.