Application of reinforcement learning to autonomous aircraft in partially observable environments
University of New Brunswick
This thesis provides a brief survey of the mathematical background of the reinforcement learning (RL) method and sketches the current state of arguably the most developed area of RL application, to the problem of controlling autonomous vehicles (self-driven car-like vehicles). This is then compared to RL solutions in autonomous piloting tasks. Contrasting the two shows that the latter may benefit from a common framework for RL applications. We propose a framework for autonomous piloting tasks, provide a detailed description of the toolkit available for the framework, and perform an experiment with described instruments. The experiment is designed to determine whether a small fixed memory window can mitigate the adverse influence of such unobserved factors as wind bursts and turbulence. Tests show that the memory mechanism that encapsulates control feedback is an informative input for the learning agent, as long as the unobserved factors affect control behavior significantly.