Figure 6. Model simulations. A, Schematic illustration of an adaptive controller: sensory feedback is used to estimate both the state of the system and the model, which in turn updates the controller and estimator online. The solid arrows represent the parametric state-feedback control loop, and the dashed arrows represent the real-time learning and update of model parameters. B, Reproduction of a standard learning experiment (hand paths and maximum lateral deviation across trials), with mirror-image catch trials (ct), and unlearning following the catch trials (difference between ct-1 and ct+1). The trial-by-trial changes were reproduced with online computations exclusively (Thoroughman and Shadmehr, 2000). C, left, Simulation of the results from experiment 1 with three distinct values of online learning rate (
γ
: 0.1 in black, 0.25 in gray, and 0.5 in dashed), which reduces the second peak hand force while the first displays smaller changes across the tested range of
γ
. The change in peak force relative to the mean across simulations was less than 2 N for the first peak, while the second peak displayed changes ∼4 N. Right, Correlations between simulated perturbation force and the force produced by the controller. Black dots are the results of the adaptive control model with tested values of
γ
(0, 0.1, 0.25, and 0.5). Gray and open dots are the results of the model with time-evolving cost-function (both increase and decrease). D, Simulations of behavior from experiment 2, the FF was turned off at the via-point (t = 0.6 s), and the second portion of the reach displays an after effect (see Materials and Methods). The inset highlights the lateral velocity measured at the minimum and at the second peak hand speed as for experimental data. The traces were simulated for distinct values of online learning rate (0.25 in dashed, and 0.5 in solid).