Dissertation
Effective Reinforcement Learning With Information Reuse From Multiple Demonstrators
Washington State University
Doctor of Philosophy (PhD), Washington State University
07/2025
DOI:
https://doi.org/10.7273/000007895
Abstract
Reinforcement learning (RL) methods may suffer from slow learning and poor initial performance in complex domains. Learning from demonstration (LfD) has emerged as a successful technique to speed up RL. This dissertation focuses on how to effectively extract and reuse information from multiple demonstrators (or demonstrations). First, we introduce the Flexible Two-level Structured Approach (FTSA), which combines action advice from multiple demonstrations using insights from contextual bandit problems and probabilistic policy reuse. Second, we develop the Two-Level Actor-Critic (TL-AC) network structure that can dynamically determine when and which demonstrator’s advice to incorporate during the learning process. Our experimental evaluations across multiple domains demonstrate that these approaches could efficiently improve learning. Additionally, we present initial investigations into comparing different advising modalities. This research contributes to the broader goal of creating sample-efficient RL systems capable of leveraging imperfect prior knowledge in complex, real-world applications.
Metrics
6 File views/ downloads
14 Record Views
Details
- Title
- Effective Reinforcement Learning With Information Reuse From Multiple Demonstrators
- Creators
- Su Zhang
- Contributors
- Matthew E Taylor (Co-Chair)Lawrence B Holder (Co-Chair)Yan Yan (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- School of Electrical Engineering and Computer Science
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Publisher
- Washington State University
- Number of pages
- 92
- Identifiers
- 99901297594001842
- Language
- English
- Resource Type
- Dissertation