查看文章

Hierarchical structure is ubiquitous in human and animal behavior: simple actions are combined to form subtask sequences, which in turn get combined into more complex(and temporally extended) tasks. Here we investigate the neural substrates of hierarchical behavior, based on the computational framework of hierarchical reinforcement learning(HRL). We posit that in order to construct this hierarchical structure, learning must occur at multiple levels at once. The HRL account requires the presence of multiple reward prediction errors, each pertaining to a different level in the hierarchy.