Erik Schlicht Department of Psychology "Statistical Decision Theory for Human Perception-Action Cycles" Thursday October 21st 12:20-1:10 Lind 409

Abstract
Statistical Decision Theory (SDT) has traditionally been used for economic modeling and game theory. More recent applications of SDT include pattern recognition, robotics, and computer vision. This talk attempts to demonstrate that SDT has utility for describing human perception-action cycles. First, the talk will provide psychophysical evidence that humans plan for sensory uncertainty when making reaching movements. To test this idea, subjects were required to repeatedly reach to the same object while gaze direction was varied, both when the target was in view and when it was occluded. The results display that observers' reach parameters systematically change as a function of the expected optimal uncertainty, even with vision obscured. Moreover, it is shown that these changes in reach behavior are predictable from a Bayesian model that quantifies the sources of uncertainty in the task: 1) visual uncertainty; 2) haptic uncertainty; and 3) sensorimotor transformation noise. The second portion of the talk overviews a natural loss function for human reach and grasp. Previous work (Trommershauser, et. al., 2004) has demonstrated that people's mean end-point location on a pointing task is such that it maximizes the expected reward for the task. However, the reward associated with each reach (i.e., the loss function) is experimentally imposed in most work of this sort. The talk will overview our work in deriving natural loss functions that may be used to predict people's actions in everyday tasks. To that end, a parametric loss function for reaching tasks was developed that is based on the physical properties of the target, the configuration of the object with respect to gravity and the observer, and the biological cost associated with the reach. A key assumption is that people plan for the object's motion when making a reach and grasp. Using this framework, predictions can be made about how people should reach if they were to minimize the expected risk for this loss function. To test the model, people were required to reach to objects at varying orientations. The results suggest that people are reaching in a manner that minimizes their expected risk for a natural loss function. The properties of this loss function and their consequences for perceptually guided reaching will be discussed.