Figure 3.

A, Relation between decreases in learning and action bias. Stimulation-related decreases in accuracy were positively correlated with an increased bias toward repeating a button press after reward trials (win-same button; Pearson's r = 0.77, p = 0.006). Each dot represents a subject, the solid black line is the regression slope, and the dashed lines represent 95% confidence intervals. B, C, The Q learning model is insufficient to explain stimulation-related behavioral changes. Simulated behavior of a standard two-parameter reinforcement learning algorithm (Q model) on a two-alternative probability learning task with inconsistent stimulus-response mapping. Accuracy (light gray line), probability of repeating rewarded items (win-stay, dark gray line), and probability of repeating rewarded actions (win-same button, black line) are shown for decreasing learning rates (α; B) and increasing noise in the choice policy (β; C). Decreases in learning rate and increases in decision noise were accompanied by a decrease in accuracy and a decrease in win-stay, but no change in win-same button.