Search Papers | Poster Sessions | All Posters
Poster A7 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Dynamic self-efficacy updating leads to optimistic overgeneralization
Jing Li1 (), Angela Radulescu2; 1Icahn School of Medicine at Mount Sinai
Humans often need to make predictions about future reward by generalizing from similar past experiences. Positive overgeneralization occurs when a rewarding experience is attributed to multiple states of the environment. In this work, we show that a model-free RL agent that dynamically updates its self-efficacy beliefs as it approaches a goal learns optimistic overgeneralized value representations. We suggest that dynamic self-efficacy beliefs may underlie the human tendency to overgeneralize from positive outcomes and test the predictions of our model in a novel behavioral paradigm designed to measure positive overgeneralization across a 1D perceptual state space. We find preliminary evidence that participants with higher self-reported general self-efficacy overgeneralize from rewarding outcomes, and place greater valuation on rewarded trials than those with lower self-efficacy beliefs, as reflected in faster reaction times.
Keywords: self-efficacy reinforcement learning positive overgeneralization computational psychiatry