Abstract
Recent models of oculomotor control have been successful at describing saccade velocity profiles using optimal control principles. Some models postulate that the eyes minimize the expected deviation from a target end point (Minimum Variance models). Other models postulate that eye movements minimize the time required to reach the target point (Minimum Duration models). These models have a common assumption that the goal of oculomotor control is to reach target points. Not surprisingly, much of the empirical data used in these models is based on tasks in which the explicit goal is for the eyes to move to predifined target end points. However, such tasks seldom occur in daily life. Instead, the eyes typically play a supportive role, providing other actuators (e.g., the hands) with the information they need to efficiently achieve their goals (e.g., grasp objects). Here, we use a rapid-pointing task to study eye movement in different conditions where the eyes serve either a supporting role (where the reward depends on the hand endpoint) or an executive role (where the reward depends on the fixation endpoint). Our results suggest that Minimum-variance and Minimum-duration models cannot account for key properties of the eye movements observed in our data. To address this issue, we present an alternative class of models (Infomax models) in which the eyes move to maximize the information needed to achieve goals. The approach relies on a novel algorithm (PIC2) developed at our laboratory to find approximate solutions to continuous time partially observable stochastic optimal control problems. We present our progress solving Infomax Control problems and show how they explain the complex saccadic movements observed in our experiments.
Meeting abstract presented at VSS 2012