Can Visual Fixations Explain Context-Dependent Reinforcement Learning?

Authors: Rishi Heggawadi, William Hayes, Melanie Touchard, Andrew Dolinsky, Nadiah Layne

SUNY Campus: Binghamton University

Presentation Type: Poster

Location: Old Union Hall

Presentation #: 23

Timeslot: Session A 9:00-10:00 AM

Abstract: Context-dependent reinforcement learning (RL) suggests that decision making is influenced by relative valuation rather than the absolute expected utility of an option. While previous studies have utilized RL models with context-dependent valuation mechanisms to understand how context shapes the decision-making process, this study aimed to test an alternative explanation based on selective visual attention to choice outcomes. Our primary goal was to test what kind of model best explains context-dependent choice biases. Fifty participants completed a learning task where they repeatedly chose between symbol pairs associated with variable rewards, followed by a transfer test involving novel symbol pairings without feedback. Participants' goal was to accumulate as much reward as possible. Using eye-tracking methods, we could record which outcomes participants fixated on throughout the learning task. Our results, across the comparison of eight different models, indicate that decisions are made based on a combination of relative and absolute values, and that visual fixation patterns alone may not provide enough information as to which types of values are learned. However, visual attention may play a bigger role in more complex environments with greater amounts of available feedback. This study highlights the importance of integrating computational modeling and eye-tracking data to better understand the underlying processes involved in reinforcement learning and decision-making.