Predicting character discrimination difficulties using gaze data while browsing the Web using reinforcement learning

Abstract

Aiming to construct a devise that magnifies characters automatically when users feel difficult to identify them while browsing websites, this study infers whether the difficulty occurs promptly and correctly based on gaze data. Precisely, having fixation duration evolve step by step, an algorithm to judge the occurrence of the difficulty in each time duration is built based on the SARSA method. As gaze fixations arise by many causes and fixation durations by difficulty to identify characters are not necessarily longer than those by the other causes, it is very difficult to determine a threshold to magnify the characters. Therefore, there are two types of errors relating to the determination of policies by the agent. One (type I error) is magnifying characters when users do not feel difficult to identify them; the other (type II error) is having the agent not magnify characters when users feel so. To solve this problem, this study tries to decrease the occurrence frequency of type I errors along with to keep that of type II errors within a permissible range by assessing precisely revenues and losses regarding the determination of policies and by utilizing not only fixation durations but other gazing data.