Abstract
Real-world tracking requires not only maintaining estimates of objects' current locations, but also extrapolating where they will move and using new information to update both state estimates and predictions. We suggest that people accomplish this efficiently by considering a limited number of potential future paths of an object that are updated only when they fail to predict how the world has unfolded. We compare humans to such a rational process model in a task using online judgments of path extrapolation during occlusion. Participants predicted whether a ball bouncing around a bumper table would reach one of two 'goals' first. Participants made continuous predictions by pressing and holding one of two buttons throughout a trial ('no prediction' was allowed). This task was performed on 400 tables, and crucially, many tables contained occlusions that hid the ball as it traveled under them. We compared these predictions to those produced by a model uses a limited number of noisy physical simulations to make predictions, and updates both the current and future states of the world without needing to generate new predictions after each update. This model accounted for whether participants would make any prediction at a given point in time (r=0.89), which goal participants were more likely to choose at each time point (r=0.83), and how often participants changed their predictions on each trial (r=0.79). It also captured how often participants switched predictions when the ball was occluded (r=0.65), suggesting that this model approximates how people store and update representations of the world even when objects are hidden. Finally, this model explained participants' predictions better than an alternate model that assumed people sample predictions from a full posterior predictive distribution. These results suggest that people use computationally efficient physics-based models to track object locations and predict how the world will unfold.
Meeting abstract presented at VSS 2014