Evolutionary Reinforcement Learning

Abstract

Recent research in bio-inspired artificial intelligence potentially provides solutions to the challenging problem of designing fault-tolerant and robust flight control systems. The current work proposes SERL, a novel Safety-informed Evolutionary Reinforcement Learning algorithm, which combines Deep Reinforcement Learning (DRL) and neuro-evolutionary mechanisms. This hybrid method optimises a diverse population of non-linear control policies through both evolutionary mechanisms and gradient-based updates. We apply it to solve the attitude tracking task on a high-fidelity non-linear fixed-wing aircraft model. Compared to a state-of-the-art DRL solution, SERL achieves better tracking performance in nine out of ten cases, remaining robust against faults, changes in initial conditions and external disturbances. Furthermore, the work shows how evolutionary mechanisms can balance performance with the smoothness of control actions, a feature relevant for bridging the gap between simulation and deployment on real flight hardware.