Reinforcement learning in artificial intelligence and neurobiology

Abstract

Reinforcement learning (RL), a computational framework rooted in behavioral psychology, enables agents to learn optimal actions through trial and error. It now powers intelligent systems across domains such as autonomous driving, robotics, and logistics, solving tasks once thought to require human cognition. As RL reshapes artificial intelligence (AI), it raises a critical question in neuroscience: does the brain learn through similar mechanisms? Growing evidence suggests it does.

To bridge this interdisciplinary gap, this review introduces core RL concepts to neuroscientists and clinicians with limited AI exposure. We outline the agent–environment interaction loop and describe key architectures including model-free, model-based, and meta-RL. We then examine how advances in deep RL have generated testable hypotheses about neural computation and behavior. In parallel, we discuss how neurobiological findings, especially the role of dopamine in encoding reward prediction errors, have inspired biologically grounded RL models. Empirical studies reveal neural correlates of RL algorithms in the basal ganglia, prefrontal cortex, and hippocampus, supporting their roles in planning, memory, and decision-making. We also highlight clinical applications, including how RL frameworks are used to model cognitive decline and psychiatric disorders, while acknowledging limitations in scaling RL to biological complexity.

Looking ahead, RL offers powerful tools for understanding brain function, guiding brain–machine interfaces, and personalizing psychiatric treatment. The convergence of RL and neuroscience offers a promising interdisciplinary lens for advancing our understanding of learning and decision-making in both artificial agents and the human brain.