Flight Control System Clearance using Bayesian Optimised Deep Reinforcement Learning on an Unstable Platform

This paper presents a novel approach for clearing Flight Control Systems (FCS) on unstable aircraft by utilising meta-optimised reinforcement learning to identify worst-case scenarios that could lead to departures from controlled flight. The authors propose a system where a Deep Reinforcement Learning (DRL) agent, using an on-policy actor-critic framework is trained to find breaches in FCS clearance requirements by interacting with a high-fidelity non-linear aircraft simulation via stick, throttle and pedal commands. To ensure the learned manoeuvres are operationally relevant, a manoeuvre monitor constrains the agent to fly realistically. Furthermore, Bayesian Optimisation is employed as an outer-loop search for the worst-case trim conditions and aircraft configurations. Results demonstrate the trained policy's ability to execute angle of attack and angle of sideslip departure manoeuvres, highlighting the potential this optimization methodology has to enhance the traditionally complex and cognitive process of FCS certification by systematically exploring high-dimensional parameter spaces.