Graduation Semester and Year
2016
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Aerospace Engineering
Department
Mechanical and Aerospace Engineering
First Advisor
Atilla Dogan
Second Advisor
Brian Huff
Abstract
Designing control systems for airship has unique challenges as compared to conventional aircraft. Highly nonlinear dynamics, different mass/inertia relations, vast uncertainties in the model parameters and underactuation are the main reasons behind this. Airship dynamics is greatly influenced by the variations in the environmental (e.g.,room temperature) and internal (e.g., helium distribution in envelope) factors that can completely change the response characteristics of the blimp and make it infeasible for a model-based controller to perform. On the other hand, a skilled RC pilot can operate the manual flight easily under these conditions. This makes LfD (learning from demonstration) and RL (reinforcement learning) techniques suitable candidates to address the issues that model-based control design fails to do. In general, LfD covers the methods that aim to learn a control policy directly from the previously provided expert demonstrations. In reinforcement learning, it is aimed to reach an optimal policy through trial and error while a reward function continuously describes whether the action taken in a specific state creates good or bad outcome. This dissertation research develops a three stage LfD/RL method which uses continuous multi-dimensional states and actions. Stages and subroutines used in the method is first explained in detail, then implemented on three simple example cases to show the performance and the convergence characteristics of exploration using discrete and continuous state-action spaces. The method is used for learning and executing 1D and 2D waypoint navigation tasks of a ground vehicle (UGV) for both simulation and hardware implementation. In order to apply the method to the motion of a low speed airship, a realistic airship flight simulator is designed by performing measurements and tests and pilot demonstrations are recorded with this simulator. Finally, the method used to learn and execute commanded position and orientation tasks demonstrated by the pilot, similar undemonstrated tasks and a case when these tasks are combined to represent a full mission. It is shown that selection of correct function approximator parameters are crucial in order to obtain satisfactory response when LfD/RL method is used.
Keywords
Airship, Reinforcement learning
Disciplines
Aerospace Engineering | Engineering | Mechanical Engineering
License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.
Recommended Citation
Daskiran, Onur, "PILOT DEMONSTRATION BASED REINFORCEMENT LEARNING WITH APPLICATION TO LOW SPEED AIRSHIP CONTROL" (2016). Mechanical and Aerospace Engineering Dissertations. 338.
https://mavmatrix.uta.edu/mechaerospace_dissertations/338
Comments
Degree granted by The University of Texas at Arlington