Fig. 4.

Example learning curves. One example simulation for each model: ${\sigma }_m^2=9$, ${\sigma }_{\eta \ast }^2=16$, α = 0.15 (for the Cashaback19 and Dhawale19 models), reward criterion = adaptive (mean), target amplitude = 4 ${\sigma }_m$. The top row shows the movement endpoint behavior that the TTC method uses to estimate exploration. Each movement endpoint is composed of an aim point, motor noise and possibly exploration. The bottom row shows the underlying aim points that are adjusted during learning. Non-successful trials are covered largely by successful trials because following failure, the aim point is not (or only slightly: Dhawale19) updated