Two-Stage Strategy to Achieve a Reinforcement Learning-Based Upset Recovery Policy for Aircraft

Cao, Huanhui; Zeng, Weifeng; Jiang, Hantao; Hu, Hao; Li, Chaoran; Lu, Wenjie; Xiong, Hao

doi:10.1109/cac53003.2021.9727381

Cited by 6 publications

(2 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

Section: Introductionmentioning

confidence: 99%

“…Previous research has acknowledged the importance of including disturbances during training [10,11]. Cao et al [11] trained their stall recovery system not only for different weather conditions, but also for different weights. Changing the weight of the aircraft can create very informative examples for the Reinforcement Learning agent to train on, for example, it can learn connections between the response of the controls and angle of attack.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Increasing predictability in the response of an AI-assisted stall recovery system in complex stall conditions by expanding the knowledge-base of AI

Koopman,

Zammit-Mangion

2024

J. Phys.: Conf. Ser.

View full text Add to dashboard Cite

The environment in the cockpit of commercial aircraft is becoming increasingly complex due to the introduction of automation systems. This complexity is especially evident when malfunctions take place, making it difficult for pilots to comprehend the interconnectedness of the systems and potentially leading to loss of control. This paper investigates a novel method for creating an Artificial Intelligence-based stall recovery assistant using Reinforcement Learning by training the agent to generate a stall and subsequently recover from it. This enables training in a large training space with a simple reward function, where the agent has the ability to develop a deep understanding of the environment. Tests show that the agent is able to recover from stall at a variety of altitudes while experiencing unreliable airspeed information originating from a blocked Pitot tube system and with a better response than all baseline agents. The results indicate that restricting AI is not always necessary and, further, that too many restrictions can lead to a system that learns only shallow features, causing it to be unreliable in unforeseen circumstances.

show abstract

Section: Introductionmentioning

confidence: 99%