Georgia Institute of Technology
Evaluating the Effects of Model Simplifications on the Transference of Policies Learned in Simulation
Abstract
dc:description.abstractBoth the military and civil worlds are being transformed by the development and deployment of unmanned systems to a wider range of scenarios. As the field of unmanned systems has grown and matured, it has continuously advanced towards increasing levels of autonomy. As an example, the cars of today have gone from "maintain this speed" to "drive on this highway." As this push towards greater levels of autonomy continues, new methods for developing policies for control of these systems are required. Recent breakthroughs in reinforcement learning hope to address this problem. The primary advantage of reinforcement learning based systems is their focus on goal driven behavior. In developing reinforcement learning based policies, there is a need for significant exploration of a system’s possible state-action space. As such, modeling and simulation has been an indispensable tool. However, transferring policies from a simulated world to the real world presents its own challenges. This work develops a method for evaluating the relative importance of different possible simplifications that can be taken during the modeling process. This relies on a sampling-based method to explore the possible simplification space of a given referent model. Experiments show that this method compares favorably to a number of baseline model development strategies and can lead to significantly simplified system models that maintain similar transference properties. Additional experiments are presented that evaluate different choices within this sampling-based strategy and the effects of imperfect referent models on the resulting evaluations.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Aerospace Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2020
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Meyer, Patrick S.
- Advisor dc:contributor.advisor
-
- Mavris, Dimitri N.
- Committee members dc:contributor.committeemember
-
- Feron, Eric
- Schrage, Daniel
- Steffens, Michael
- Mariottini, Gian Luca
Subjects
dc:subject × 3Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- http://hdl.handle.net/1853/64103