Neural Computations Underlying Arbitration between Model-Based and Model-free Learning

摘要

There is accumulating neural evidence to support theudexistence of two distinct systems for guiding actionudselection, a deliberative "model-based" and a reflexiveud"model-free" system. However, little is knownudabout how the brain determines which of theseudsystems controls behavior at one moment in time.udWe provide evidence for an arbitration mechanismudthat allocates the degree of control over behaviorudby model-based and model-free systems as a functionudof the reliability of their respective predictions.udWe show that the inferior lateral prefrontal and frontopolarudcortex encode both reliability signals andudthe output of a comparison between those signals,udimplicating these regions in the arbitration process.udMoreover, connectivity between these regions andudmodel-free valuation areas is negatively modulatedudby the degree of model-based control in the arbitrator,udsuggesting that arbitration may work throughudmodulation of the model-free valuation systemudwhen the arbitrator deems that the model-basedudsystem should drive behavior.