Human-Centered Specification and Explanation for Mixed-Initiative Interactions
Author(s)
Tambwekar, Pradyumna
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
As Artificial Intelligence systems become commonplace in society, we require bidirectional, human-centered, mechanisms of communication.
Particularly in a mixed-initiative setting, wherein humans and AI systems are collaborating towards a shared goal, humans should be able to specify their intentions to AI-systems and also interpret the intentions of an AI-collaborator.
Ill-fitting methods can undermine the human-AI interaction, leading to a downgrade in performance and an increase in mistrust of autonomous systems.
In this thesis, I build and study human-centered methodologies which enable humans to specify their desired behavior of an AI system, as well as receive suitable explanations which optimize their ability to perform the shared task.
First, I build two machine learning frameworks to enable humans to specify their desired behavior of an autonomous system via unstructured natural language.
In my first contribution, I develop a machine learning framework that translates a user's unstructured description of their desired policy into a decision-tree, which is then utilized to initialize a differentiable decision tree (DDT) policy. The use of a ``white-box'' framework such as a DDT, enables users to interpret the final learned policy of the agent.
Next, I expand this method towards interpreting the high-level strategies of a user rather than a description of a specific policy.
For this task, I train a computational interface, powered by a large-language-model, to translate language descriptions of strategic intent into actionable ``Commander's Intent'' in the form of goals and constraints.
The second half of my thesis pertains to explainable-AI, specifically, understanding the factors which influence a user's interaction with an explanation and developing personalizable explanation methods which consider the specific user and the context of the interaction.
First, I conduct an human-subjects experiment aimed at understanding the differences, wherein we observe a counter-intuitive phenomenon; participants performed significantly worse with explanations they preferred at a significant level.
To address this finding, I developed a first-of-its-kind personalized explainable AI framework to adaptively balance a participants preference and task-performance.
This method is comprised of two components. First, I build a federated, personalization model which predicts when a participant is likely to make a mistake. Second, I utilize this prediction to provide a suitable explanation to a user, i.e. if a user is likely to be incorrect, they are more likely to be given an explanation that optimizes their performance and if they are likely to be correct, they are more likely to be given an explanation which they prefer.
Finally, I conclude this thesis with a human-subjects experiment which compares different forms of personalization for explainable AI.
We compare adaptive personalization with adaptable personalization to study the impact of varying the approach to personalizing an explanation received by a user.
Sponsor
Date
2024-08-27
Extent
Resource Type
Text
Resource Subtype
Dissertation (PhD)