Authors:

Matthew Beecher

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Automation is increasingly used for improving system performance across a broad range of applications. For example, intelligent decision aid systems, designed to aid humans in a variety of tasks, such as navigation. In such systems, it is vital the operator appropriately utilize the decision aid without over or under-reliance. In other words, the operator's trust should be calibrated to the reliability of the decision aid. In prior research, researchers showed that a partially observable Markov decision process (POMDP) model could be used to estimate an operator's trust state by observing their behavior, namely decision-aid reliance. The POMDP was then used to design a near-optimal policy to adapt the user interface transparency displayed by the decision aid to operators' trust to mitigate the consequences of improper reliance. However, this work was limited in that a single POMDP was trained using an aggregated dataset of human data; additional research has identified the presence of two distinct trusting behaviors among the dataset, with two distinct POMDP models trained and associated adaptive transparency policies designed and synthesized. This work aims to implement the two proposed policies in closed-loop (conducted via Amazon Mechanical Turk) and evaluate their efficacy in calibrating trust as compared to the generalized policy. It will utilize a classifier trained to identify operator trust behavior with behavioral data during the experiment to then determine which control policy the operator experiences. This research will illustrate the usefulness of this split modelling approach in a practical setting and open the door for further developments.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Matthew Beecher