Files and links (1)
CC BY V4.0, Open Access
Thesis
The Impact of Different Summaries as Reinforcement Learning Explanations on Human Performance And Perception
Washington State University
Master of Science (MS), Washington State University
01/2021
DOI:
https://doi.org/10.7273/000002416
Handle:
https://hdl.handle.net/2376/121035
Abstract
As deep reinforcement learning’s capabilities surpass those of traditional tabular reinforcement learning, the reinforcement learning community is working to make these algorithms less opaque. Explanations about the algorithms’ choices and strategies serve this purpose. However, deep reinforcement learning algorithms as they exist today do not make information about their operation easily accessible, making explanations difficult to build. The research in this thesis aimed to extract such information, use it to build explanations, and then test those explanations in a user study.
Eight different deep reinforcement learning agents were trained using the OpenAI Baselines implementation of DQN. Then, the HIGHLIGHTS-DIV algorithm was altered to both create video-based summaries of the agents’ strategies and to collect a broad range of data about the agents’ interactions with the game Ms. Pacman. The collected data was used to create additional explanations which were then compared to the video-based summaries in a user study. The between-subjects user study had 232 participants answer questions about two different agents, selected from a set of four agents. To answer the questions, participants were given either a video-based summary, the alternative data-based explanation, or a combination. The questions included predicting the agents’ next moves, identifying regions of the environment important to the agents, identifying risks to the agent, and assessing the agents’ overall capabilities.
The results of the user study revealed that the alternative explanations built from the collected data led to more correct answers about the agents and their strategies. Additionally, the study showed that different explanations may be more or less helpful, depending on the context. Finally, the study showed that users’ reported trust in an assessment does not directly correlate to performance. This strongly suggests that trust and effectiveness should be measured and calibrated separately in future examinations of explanations in deep reinforcement learning.
Metrics
5 File views/ downloads
28 Record Views
Details
- Title
- The Impact of Different Summaries as Reinforcement Learning Explanations on Human Performance And Perception
- Creators
- Brittany Faith Davis Pierson
- Contributors
- John Miller (Advisor)Dustin Arendt (Committee Member)Matthew E Taylor (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Engineering and Applied Sciences (TRIC), School of
- Theses and Dissertations
- Master of Science (MS), Washington State University
- Publisher
- Washington State University
- Number of pages
- 238
- Identifiers
- 99900606753701842
- Language
- English
- Resource Type
- Thesis