The structure of reinforcement-learning mechanisms in the human brain

- Reinforcement-learning algorithms are a useful framework for investigating the neurobiology of action selection for reward.
- Multiple approaches have been taken in the literature, including the distinction between model-based and model-free RL, hierarchical reinforcement-learning and state-space or structure learning.
- There is evidence for both model-based and model-free RL signals in the human striatum, parietal and frontal cortices and elsewhere.
- Arbitration between model-based and model-free RL mechanisms may involve regions of lateral and frontopolar cortex.
- Each of these approaches focus on different problems in neural RL, and an outstanding challenge is to find a way to integrate them.

Here we review recent developments in the application of reinforcement-learning theory as a means of understanding how the brain learns to select actions to maximize future reward, with a focus on human neuroimaging studies. We evaluate evidence for the distinction between model-based and model-free reinforcement-learning and their arbitration, and consider hierarchical reinforcement-learning schemes and structure learning. Finally we discuss the possibility of integrating across these different domains as a means of gaining a more complete understanding of how it is the brain learns from reinforcement.