Prediction of user action in moving-target selection tasks
Selection of moving targets is a common task in human–computer interaction (HCI), and more specifically in virtual reality (VR). In spite of the increased number of applications involving moving–target selection, HCI and VR studies have largely focused on static-target selection. Compared to its static-target counterpart, however, moving-target selection poses special challenges, including the need to continuously and simultaneously track the target and plan to reach for it, which may be difficult depending on the user’s reactiveness and the target’s movement. Action prediction has proven to be the most comprehensive enhancement to address moving-target selection challenges. Current predictive techniques, however, heavily rely on continuous tracking of user actions, without considering the possibility that target-reaching actions may have a dominant pre-programmed component—this theory is known as the pre-programmed control theory.
Thus, based on the pre-programmed control theory, this research explores the possibility of predicting moving-target selection prior to action execution. Specifically, three levels of action prediction are investigated: action performance, prospective action difficulty, and intention. The proposed performance models predict the movement time (MT) required to reach for a moving target in 2-D and 3-D space, and are useful to compare users and interfaces objectively. The prospective difficulty (PD) models predict the subjective effort required to reach for a moving target, without actually executing the action, and can therefore be measured when performance can not. Finally, the intention models predict the target that the user plans to select, and can therefore be used to facilitate the selection of the intended target.
Intention prediction models are developed using decision trees and scoring functions, and evaluated in two VR studies: the first investigates undirected selection (i.e., tasks in which the users are free to select an object among multiple others), and the second directed selection (i.e., the more common experimental task in which users are instructed to select a specific object). PD models for 1-D, and 2-D moving-target selection tasks are developed based on Fitts’ Law, and evaluated in an online experiment. Finally, MT models with the same structural form of the aforementioned PD models are evaluated in a 3-D moving-target selection experiment deployed in VR. Aside from intention predictions on directed selection, all of the explored models yield relatively high accuracies—up to ~78% predicting intended targets in undirected tasks, R^2 = .97 predicting PD, and R^2 = .93 predicting MT.