Authors:

Douglas Nyberg

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Technological advances in autonomous agents are demonstrating their capabilities as effective decision aids in humans' daily activities. For decision aids to be fully effective, issues of human overtrust and undertrust in autonomous recommendations must be addressed. It has been shown that user interface (UI) transparency in autonomous agent recommendations affect human trust. Ideally, each individual would have a tailored trust model, but this is currently unrealistic due to the required computational resources and the practicality of obtaining individualized data. Therefore, an optimal comprise is to group individuals into defined trust groups in order to balance model efficiency and computational requirements. Previous research has identified two distinct groups for optimal trust balancing using Partially Observable Markov Decision Process (POMDP) Models. In this work, we utilize an ensemble learning method comprised of logistic regression, a random forest classifier, and a support vector machine to determine into which group individuals should be sorted. By implementing recursive feature elimination methods, features generally highlighted for training the classifier were the humans' compliance, responses, trial times, and response times to the decision aid's recommendations. With trust groups predicted, transparency can now be altered according to a control policy that has been designed based on that individual's dominant trusting behavior, or trust group. Finally, a closed- loop human user study is conducted to determine if the classification method is effective for determining optimal UI transparency, thereby improving the calibration of users' trust and their reliance behavior with respect to the intelligent decision aid.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Douglas Nyberg