Biological reward codes in computational reinforcement learning

What is it about?

Novel, uncertain, or dynamic environments require organisms to learn appropriate behavior based on environmental feedback, a process widely modeled with reinforcement learning (RL) algorithms. Despite their prominence and performance, standard RL models assume linear reward functions at odds with empirical nonlinearities in neural reward representations. Here we analyze a nonlinear RL algorithm implemented via the canonical divisive normalization computation, and show that this model introduces a tunable asymmetry in reward learning that captures known diversity in both behavioral and neural responses. These findings suggest that biological value functions confer adaptive flexibility in reward-guided learning and decision-making.

Featured Image

Photo by David Clode on Unsplash

Photo by David Clode on Unsplash

Why is it important?

Reinforcement learning (RL) is a fundamental pillar of artificial intelligence and machine learning, widely applied in computer science, psychology, and neuroscience. Here, we show that combining RL algorithms with a biological reward code provides computational flexibility in both neural information processing and behavior. These findings offer insight into how the brain learns and represents rewards and suggests ways to improve computational RL algorithms.