(A) In the changepoint condition, surprising events (changepoints) signaled a transition in the aim of the cannon whereas (B) in the oddball condition, surprising events (oddballs) were unrelated to the process through which the aim of the cannon transitioned. (C) Learning rate in the cannon task can be described by the slope of the relationship between prediction error (signed distance between cannonball and shield; abscissa) and update (signed change in shield position after observing new cannonball location; ordinate). Fixed learning rate updating corresponds to a line in this space whose slope is uniform across prediction errors and reflects the learning rate (gray lines). In contrast, normative learning dictates that the slope should decrease for extreme prediction errors in the oddball condition (blue) but increase for extreme prediction errors in the changepoint condition (orange). (D) Prediction error (abscissa) and update (ordinate) for each trial (points) in each condition (designated by color) completed by a single example participant. Size of points is inversely related to density of data for improved visualization. (E) Trial updates for each subject were fit with a regression model that included prediction errors (to measure fixed learning rate) as well as several interaction terms to assess how learning depended on various factors. (F) Coefficients from regression model fit to individual subjects (points) revealed an overall tendency to update toward recent cannonball locations (red, t = 14.4, dof = 38, p=10−17), and a tendency to do so more in the changepoint condition (green, t = 3.1, dof = 38, p=0.003), when uncertain (yellow, t = 7.5, dof = 38, p=4×10−9), and on trials where the cannonball was not blocked by the shield (pink, t = −3.4, dof = 38, p=0.001). The model revealed that there was no consistent effect of surprise on learning across both conditions (blue, t = 0.8, dof = 38, p=0.43), but that there was a strong interaction between surprise and condition (orange, t = 9.9, dof = 38, p=4×10−12) whereby surprise tended to increase learning in the changepoint condition but decrease learning in the oddball condition.