Striosome-dopamine circuit signals information gain, not prediction error

Abstract The striosome compartments of striatum gate cortical signals to dopamine cells, which in turn project back to striosomes and the surrounding matrix¹. This loop is essential for decision-making, learning, and movement²,³. This circuit’s dopamine has long been thought to signal reward prediction errors (RPEs)e.g. 4,5. But RPEs struggle to account for dopaminergic responses to large costs, novelty, or moment-to-moment effect on movement and decisions⁶–¹⁰. Here we demonstrate that this circuit’s dopamine signal instead computes policy information gain (policy-IG), a single information-theory measure of the benefit of new information for refining present and future action strategies. For example, how much does a weather report increase probability of packing an umbrella? We show that policy-IG reduces to classic RPEs in simple reward tasks. However, policy-IG also predicts dopaminergic responses to aversive stimuli⁷,⁸, nonlinear relationship with reward levele.g. 11, response to novel stimuli⁷, role in learning causal associations¹², correlation with state value¹³, and moment-to-moment guidance of movement¹³ and decisions¹⁴–¹⁶. Further, impairing policy-IG in our model reproduces key features of disorders affecting the basal ganglia and sheds light on the development of neuropsychiatric disorders. These results identify policy-IG as a promising target for therapies that restore balanced decision-making and for artificial intelligence algorithms modeled on this circuit.