The Role of Eye-Movements in the Representation of Value in a Reinforcement Learning Context: A Web-Based Study

Abstract

The present thesis experimentally set out to try to answer if there was a correlation between reinforcement learning and eye-movements and what the implications of such a correlation might be. An important experimental factor here was the decision to do this online, to see if it was possible to get valid and reliable results, and furthermore perhaps reach out to a more diverse group of people than a typical in-lab study would, making the results more generalizable. 38 people were recruited via the website prolific.co. The participants then performed a learning test, where they were shown two symbols on the screen, and the objective was to find the symbol with the highest value out of the pair, with three different symbol pairs. After each presentation of a pair, the participant had to choose one symbol, which were then followed by a rewarding or non-rewarding feedback. Each symbol pair had a different ratio of positive relative to negative feedback. The participants eye-movements were tracked via their web-camera, to see if they fixated more on the most rewarding symbol. The results showed that the participants reliably learned to choose the higher value symbol, comparable in validity to that of in-lab studies. We also found a statistically significant correlation between learning and fixating on the most rewarding symbol, although the quality of the eye-tracking was of too poor quality to draw any conclusions about this correlation. The present experiment reached a diverse group of people from all over the world and proved that it is possible to perform a reinforcement learning experiment online, although the technology of eye-tracking cannot match an in-lab study. Further online research is needed in many areas to determine what type of experiments can produce valid and reliable data, which is especially relevant to the generalizability of research, and the present situation of a global pandemic which limits the in-lab approach.