상세정보
Export to Refworks부가기능
Reinforcement Learning with High-Level Task Specifications
상세 프로파일
| 자료유형 | 학위논문 |
|---|---|
| 서명/저자사항 | Reinforcement Learning with High-Level Task Specifications. |
| 개인저자 | Wen, Min. |
| 단체저자명 | University of Pennsylvania. Electrical and Systems Engineering. |
| 발행사항 | [S.l.]: University of Pennsylvania., 2019. |
| 발행사항 | Ann Arbor: ProQuest Dissertations & Theses, 2019. |
| 형태사항 | 172 p. |
| 기본자료 저록 | Dissertations Abstracts International 81-03B. Dissertation Abstract International |
| ISBN | 9781088366899 |
| 학위논문주기 | Thesis (Ph.D.)--University of Pennsylvania, 2019. |
| 일반주기 |
Source: Dissertations Abstracts International, Volume: 81-03, Section: B.
Advisor: Topcu, Ufuk |
| 이용제한사항 | This item must not be sold to any third party vendors. |
| 요약 | Reinforcement learning (RL) has been widely used, for example, in robotics, recommendation systems, and financial services. Existing RL algorithms typically optimize reward-based surrogates rather than the task performance itself. Therefore, they suffer from several shortcomings in providing guarantees for the task performance of the learned policies: An optimal policy for a surrogate objective may not have optimal task performance. A reward function that helps achieve satisfactory task performance in one environment may not transfer well to another environment. RL algorithms tackle nonlinear and nonconvex optimization problems and may, in general, not able to find globally optimal policies. The goal of this dissertation is to develop RL algorithms that explicitly account for formal high-level task specifications and equip the learned policies with provable guarantees for the satisfaction of these specifications. The resulting RL and inverse RL algorithms utilize multiple representations of task specifications, including conventional reward functions, expert demonstrations, temporal logic formulas, trajectory-based constraint functions as well as their combinations. These algorithms offer several promising capabilities. First, they automatically generate a memory transition system, which is critical for tasks that cannot be implemented by memoryless policies. Second, the formal specifications can act as reliable performance criteria for the learned policies despite the quality of the designed reward functions and variations in the underlying environments. Third, the algorithms enable online RL that never violates critical task and safety requirements, even during exploration. |
| 일반주제명 | Artificial intelligence. Computer science. |
| 언어 | 영어 |
| 바로가기 | 
: 이 자료의 원문은 한국교육학술정보원에서 제공합니다. |
