Reinforcement learning with multiple representations in the basal ganglia loops for sequential motor control

The basal ganglia (BG) have been hypothesized to perform reinforcement learning by use of reinforcement signals provided by dopamine neurons. It is well known that there exist multiple BG-thalamocortical loops, but their functions are poorly understood. Here, the authors propose a computational model of how different BG loops are employed in visuomotor sequence learning using different representations of sequence. The central idea of the model is that a visuomotor sequence is easier to learn in spatial representation (e.g. visual coordinates) but is easier to control in body-based representation (e.g. joint angle coordinates). The results of simulations of the model replicated both behavioral and neurophysiological findings in experimental studies using a “2×5 task”