What is a reward prediction error?

This broader definition of error differs from the common sense meaning of error as having done something wrong (which also constitutes a discrepancy, namely what I have done versus what I should have done). A reward prediction error, then, is the difference between a reward that is being received and the reward that is predicted to be received.

Does the reward prediction error switch internal models?

Nevertheless, we found that the reward prediction error plays a significant role in model switching in the human brain. Thus, this result contradicts the previous theory. However, our simulation with the policy gradient shows that the reward prediction error alone can efficiently switch internal models.

Is the Bayesian estimation driven by the reward prediction error falsifiable?

In addition, if we incorporate the reward prediction error into a part of the state variable, the computational model with Bayesian estimation driven by the reward prediction error is also available (Singh, 1992, Sugimoto et al., 2012) (see Supplemental discussion S1). In this sense, these two models do not satisfy falsifiability.

What are the different types of prediction errors?

Computationally, two different types of prediction errors are observable to the human brain: reward prediction errors and sensory prediction errors.