Abstract
Purpose. Observers' choice of saccade endpoint in searching for a target embedded in noise is well modeled by an ideal observer (Najemnik & Geisler, Nature, 2005). In this study we ask whether observers' choices of saccade endpoint optimally take rewards into account. Method. Observers searched for a target in noise. On each trial, the observer fixated the center of a display. Eight Gaussian-white-noise disks (1.5 deg diam, 10 deg eccentricity) appeared, equally spaced around fixation, to one of which was added a low-contrast Gabor-patch target. Correct detections in most locations resulted in a 100-point reward, and either the top (90 deg) or bottom (270 deg) position had a reward of 500 points (indicated in advance). There was no reward for incorrect responses. The contrast of the Gabor patch was adjusted so that d' = 1 at 10 deg eccentricity. The observer had 250 ms to initiate a saccade. The target remained visible for 200 ms following the end of the saccade, thus affording a second look at the stimulus. Subjects judged which patch contained the target. A full visibility map was obtained and used to model saccadic choices of an ideal observer that maximizes expected gain. Results and Conclusions. There were significant differences between actual and ideal distributions of saccade location in all conditions (8 target positions x 2 reward positions). The ideal observer tends to make short saccades halfway between the initial fixation position and the target or high-reward position. Human observers make longer saccades, landing on or near the 8 patches. Most saccades land on or near the target or the high-reward position. For each condition, efficiency was calculated as the ratio of the human observer's actual gain to that expected of the ideal observer. Efficiencies were near optimal for targets adjacent to the high-reward location.