Abstract
Humans move their eyes multiple times every second and behind every movement is a decision, where this movement should go. Past research has predominantly focused on quantifying how properties of the task, of the scene context, or of the stimulus influence this decision. By contrast, the influence of subjective preferences on this decision has rarely been studied. One reason might be, that a fundamental problem in studying such preferences with commonly employed tasks investigating gaze shifts, is that the empirical gaze statistics are a product of all these diverse influences. Here, we introduce an experiment in the spirit of preference elicitation paradigms in economics, in which subjects reveal their subjective preferences by repeatedly deciding between alternatives. Subjects were instructed to choose between two alternative saccadic targets on each trial. We quantified individuals’ choices in terms of the saccadic amplitude, the absolute direction in visual space, and the relative change in direction relative to the previous gaze shift. All subjects showed an approximately linear preference for shorter saccades in line with known oculomotor biases also predicted by optimal motor control. Secondly, all participants showed preferences for return saccades but to varying degrees. By contrast, idiosyncratic preferences for absolute directions in visual space varied heavily between participants. All individual preferences were highly consistent throughout the experiment. To quantify and understand our subjects’ behavior, we inferred the parameters describing gaze preferences using a random utility model. Individual subjects’ choices can be described quantitatively with the relative contributions of the three features describing gaze target alternatives. This model was able to correctly predict on average 80% of participants’ decisions and the predicted utility values match the empirically observed preferences. Taken together, the experiment reveals individual differences and commonalities in oculomotor preferences and the computational model allows incorporating these preferences in gaze target prediction models.