Abstract
<p>Who we choose to befriend can impact many of our own life outcomes, from social capital to wellbeing. Here we find that reinforcement learning principles can characterize friendship decision-making and predict social connection and memory for a potential friend. Participants completed a novel friendship decision-making paradigm in which they observed video clips from the reality TV show, The Valley, depicting a clique of friends navigating friendship with a specific character, Kristen. The clips were normed to vary in the reward value of friendship with Kristen. For each scene, participants decided whether they would want to be friends with Kristen and how connected they felt to her. One day later, they completed surprise memory tests and individual-difference measures. The reward, expected value, and prediction error of a scene about Kristen predicted participants’ feelings of connection to her. Individual differences relevant to friendship formation—concern over relationship acceptance and social support receipt—corresponded with differences in exploratory versus exploitative friendship decision-making (inverse temperature) and flexibility in learning about Kristen (learning rate). Finally, memory showed a negativity bias. Participants more accurately recalled scenes characterized by lower friendship reward and larger negative friendship prediction errors. Together, these findings suggest that friendship formation can be viewed as a learning problem: people continuously update expectations about friendship with others, and those updates shape both social connection and memory. Computational principles that help us learn about rewards in the world may also help shape who we fold into our social world.</p>