The striosome compartments in the striatum gate cortical inputs to dopamine neurons, which then feed back to the striosomes and surrounding matrix. This loop underlies decision-making, learning, and movement. Dopamine signals strongly correlate with reward prediction errors (RPEs) in certain tasks. However, many dopaminergic responses, such as to high costs, novelty, aversive stimuli, and real-time movement guidance, do not align with RPE. Separately, information theory explains how dopamine responds to uncertainty and encodes when rewards are expected to occur. Here, we show that apparent RPE correlations arise mathematically from the information gain of the policy of actions (policy-IG). Policy-IG quantifies how much newly arriving information changes choice. In simple reward tasks, policy-IG reduces to classic RPEs, but it also predicts dopamine responses to aversive events, nonlinear reward scaling, novelty, movement, state valuation, and moment-by-moment decision control. Thus, RPE could be a special case of this more general function. We show that policy-IG is mathematically related to a range of information-centric, Bayesian, active inference, and casual association models, allowing those models to formally incorporate the experimental RPE literature in their support. Simulating impaired policy-IG replicates basal ganglia disorder features, suggesting policy-IG as a target for therapies. Striosomes and the brain’s dopamine system influence learning, movement, and decision-making, but it is difficult to account for these roles together. Here the authors show that dopamine signals can be understood as tracking how much new information changes what action should be taken next, unifying its diverse functions under a single principle.
Beck et al. (Sat,) studied this question.