Analysis of eye movements during relative time-to-contact judgments of younger and older observers for the development of a reinforcement learning model
Manolo Meza-Arroyo · ThinkTech (Texas Tech University) · 2015
Eye movements are essential to acquire relevant information during time-to-contact judgments. However, the relevance of such information can change while colliding objects approach, and reliably perceiving a potential collision depends on the spatiotemporal locus of fixation. In addition, this ability to effectively perceive collisions is affected by age. Current models of visual attention and eye movements do not integrate the changes in relevant information during time-to-contact judgments, and do not account for age related differences in perceptual judgments of time-to-contact. This study introduces a model based on reinforcement learning (RL) to reproduce and predict eye movements on a moment-by-moment basis during time-to-contact judgments. The present study employed a novel method in which eye movements, recorded from observers who moved a joystick to continuously report which of two approaching objects would reach them first, were used to identify visual strategies for TTC judgments. The occurrence and time of TTC responses (joystick movement), and the time and location of fixation locus were identified. The majority of observers changed their response during the approach event. Responses made during early moments when the objects were far from the observer appeared to be influenced by optical size, and responses made late in the approach event seemed to be based on expansion rate. Further analysis revealed that the distribution of TTC responses highly correlates with the η function. Younger observers demonstrated a greater proportion of correct time-to-contact judgments. Different visual strategies for younger and older observers who correctly detected the first-arriving object were found. In general, younger observers fixated significantly more time on the first-arriving sphere, while older observers fixated significantly more time on the more salient sphere (the last-arriving one). These results provided plausible explanations for the differences in correct time-to-contact judgments between age groups. Analyses of the distributions of fixation onset revealed that the frequency of fixations on the first-arriving sphere remained larger than that of the last-arriving sphere on final moments of the approach event. In addition, fixation-response sequences revealed that observers relied more on foveal vision to judge time-to-contact, and that responses relying more on peripheral vision were more likely to incorrectly judge the first-arriving object. Finally, the results obtained in this experiment were used to set different parameters in the proposed RL model. The η function found was used as a reward mechanism for eye movements. And, the distribution of fixations from younger and older observers was used to set the probabilities of the policy for executing actions (i.e., saccades and fixations). This model was able to learn the better actions for any given state.