Abstract
The reinforcement learning framework provides a computational and behavioral foundation for understanding how agents learn to maximize rewards and minimize punishments through interaction with their environment. This framework has been widely applied across disciplines, including artificial intelligence, animal psychology, and economics. Over the last decade, a growing body of research has shown that human reinforcement learning often deviates from normative standards, exhibiting systematic biases. The first aim of this paper is to propose a conceptual framework and a taxonomy for evaluating computational biases within reinforcement learning. We specifically propose a distinction between praxic biases, characterized by a mismatch between internal representations and selected actions, and epistemic biases, characterized by a mismatch between past experiences and internal representations. Building on this foundation, we characterize and discuss two primary types of epistemic biases: relative valuation and biased update. We describe their behavioral signatures and discuss their potential adaptive roles. Finally, we eleborate on how these findings may shape future developments in both theoretical and applied domains. Notably, despite being widely used in clinical and educational settings, reinforcement-based interventions have been comparatively neglected in the domains of behavioral public policy and decision-making improvement, particularly when compared to more popular approaches such as nudges and boosts. In this review, we offer an explanation for this comparative neglect that we believe rooted in common historical and epistemological misconceptions, and advocate for a greater integration of reinforcement learning into the design of behavioral public policy.