Rethinking reinforcement learning: the interoceptive origins of reward

We all love a good ice-cream. But what exactly is rewarding about consuming it? In conventional reinforcement learning models, the environment emits scalar ‘ground-truth’ reward signals that the agent can use to learn what to do. But in biological agents, ‘reward’ is subjective, dynamic and state-dependent – generated within the organism, and inferred from noisy interoceptive signals. In the Cognitive Modeling group, we study the subjectivity and flexibility of reward functions in biological agents using both experiments and conceptual work extending conventional RL models. We are also interested in how this perspective can inform our understanding of disturbances in reward learning across mental health.