Item

Abstract

People weigh positive outcomes more heavily than negative ones: a phenomenon commonly referred to as optimism bias. We hypothesized that the optimism bias can emerge as a rational strategy in light of people's experiences. To investigate this, we trained a meta-reinforcement learning agent on an instrumental learning task that elicits the optimism bias. Meta-reinforcement learning agents are known to converge to rational solutions even in complex environments. We analyzed the behavior of converged agents using standard models of reinforcement learning. The results showed that meta-reinforcement learning agents exhibit an optimism bias similar to the one observed in humans. Our findings, therefore, suggest that the optimism bias may be a rational response to people's environments.